Internet-Draft PACE November 2026
Nandakumar & Jennings Expires 20 May 2027 [Page]
Workgroup:
Agent Communication Protocols
Internet-Draft:
draft-nandakumar-agentproto-pace-01
Published:
Intended Status:
Standards Track
Expires:
Authors:
S. Nandakumar
Cisco
C. Jennings
Cisco Systems

PACE: Protocol for Agent Communication Exchange

Abstract

This document defines the Protocol for Agent Communication Exchange (PACE), an agentic dialog management protocol with a binding to Media over QUIC Transport (MOQT). A dialog is the correlated sequence of interactions between two or more participants over the course of a task. PACE defines a transport-independent dialog context and control operations to establish, modify, and terminate dialogs and to revoke propagation relationships. It specifies how that context stays coherent across intermediaries, trust boundaries, modality changes, and interruptions, and how it is carried over MOQT. PACE does not interpret or define dialog content. Over MOQT, a dialog can combine text, audio, video, and request/response exchanges in point-to-point, point-to-multipoint, and multi-party topologies. Application-layer agent protocols such as A2A, MCP, and AutoGen keep their own message formats and carry the dialog context through their own extension points. Delegation between agents and approval of individual agent actions are specified in a companion document.

About This Document

This note is to be removed before publishing as an RFC.

The latest revision of this draft can be found at https://suhasHere.github.io/draft-nandakumar-agentproto-pace/draft-nandakumar-agentproto-pace.html. Status information for this document may be found at https://datatracker.ietf.org/doc/draft-nandakumar-agentproto-pace/.

Discussion of this document takes place on the Agent Communication Protocols Working Group mailing list (mailto:agentproto@ietf.org), which is archived at https://mailarchive.ietf.org/arch/browse/agentproto/. Subscribe at https://www.ietf.org/mailman/listinfo/agentproto/.

Source for this draft and an issue tracker can be found at https://github.com/suhasHere/draft-nandakumar-agentproto-pace.

Status of This Memo

This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.

Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at https://datatracker.ietf.org/drafts/current/.

Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."

This Internet-Draft will expire on 20 May 2027.

▲

Table of Contents

1. Introduction

AI agents are autonomous software systems that work with users, other agents, and tools to complete tasks. The exchanges between them form dialogs. A dialog is the correlated sequence of interactions between two or more participants over the course of a task. For dialogs to work across platforms and vendors, participants need a common way to correlate a dialog and maintain its lifecycle across multiple hops, intermediaries, trust boundaries, and changes of modality or network attachment, without sharing the content of the dialog itself.

This document defines the Protocol for Agent Communication Exchange (PACE). PACE has two parts:

In terms of the Agent Communication Protocol architecture [ARCH], the first part realizes the Agentic Dialog Management block (Section 6.1 of [ARCH]), and the second is the MOQT profile corresponding to Section 9.1 of [ARCH].

The dialog context consists of identifiers and lifecycle state only. Content exchanged with AI models, such as conversation memory, retrieved documents, or prompts, is not part of it. The content is opaque to PACE, the MOQT binding carries it on content tracks without interpreting it. PACE does not replace application-layer agent protocols. A2A, MCP, AutoGen, and similar protocols keep their native messages and carry the dialog context through their own extension points (Section 8).

1.1. Scope and Goals

Agent interactions are increasingly live dialogs rather than one-shot requests. A user speaks while an agent answers, cuts in, and expects the agent to stop. Voice and video run alongside chat, streaming model output, and tool calls. Several agents and people work on the same task, and the task can outlast a single connection or network. In natural spoken conversation, people take turns within a few hundred milliseconds [STIVERS], so an interruption has to take effect ahead of output that is already queued. PACE is designed for these dialogs.

PACE addresses the following goals.

  • Dialog correlation: A dialog identifier that stays stable across hops, intermediaries, trust boundaries, modality changes, and time. Defined by the dialog context: identifier, generation, state, controller, and participants (Section 3.2).

  • Multi-modal exchange: Real-time media, chat, streaming model output, and request/response operations in one dialog. Each modality keeps its own latency and reliability, and none of them delays the others, even when they share one transport session (Section 5).

  • Multi-party exchange: Point-to-point, point-to-multipoint, and multi-party dialogs among agents and people. Relays fan out each message, so a sender publishes once however many participants there are, and late joiners catch up from relay caches (Section 6).

  • Low-latency interaction: Dialog control, including interruption and cancellation, takes effect within the latency of natural conversation and ahead of queued content, so that agents can hold an interactive voice conversation (Section 5).

  • Dialog lifecycle: Opening, updating, and terminating dialogs, and revoking propagation relationships, with the controller as the single authoritative source of state (Section 3.3).

  • Resilience: Recovery from path and intermediary failures, resumption after interruption, and defined behavior when a participant or the controller is unreachable (Section 3.9, Section 4.3.5).

  • Transport independence: A dialog model that any transport can carry. This document specifies a MOQT binding: namespaces, tracks, operation mapping, and delivery (Section 4).

  • Application protocol reuse: MCP, the Personal Agent Protocol, AutoGen, and A2A keep their native messages and carry the dialog context (Section 8).

  • Identity and authorization reuse: Agent identity separate from user identity, delegated authorization, and selective disclosure, without mandating a token format (Section 7, Section 9).

  • Privacy: Mitigations for linkage through persistent dialog and participant identifiers (Section 9.3).

PACE does not define trust frameworks, authorization protocols, or audit data structures; deployments use existing mechanisms, and auditing systems can use dialog identifiers. The authority agents exercise in a dialog, delegating work and approving single actions, is specified in PACE Authorization [PACE-AUTHZ]. Dialog content, model behavior, planning, and task semantics are also out of scope.

1.2. Requirements Language

The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 [RFC2119] [RFC8174] when, and only when, they appear in all capitals, as shown here.

1.3. Terminology

Dialog:

The correlated sequence of interactions between two or more participants over the course of a task. A dialog can outlast any single message exchange, MOQT session, or transport connection. A two-party dialog is an agentic dialog [ARCH], which composes larger topologies by cascading two-party dialogs; PACE supports both cascaded and multi-party dialogs (Section 6).

Dialog context:

The identifiers and lifecycle state needed to correlate and maintain a dialog (Section 3.2). The dialog context does not include content exchanged with AI models, such as conversation memory, retrieved documents, prompts, or media payloads. It corresponds to the meta-data of the [ARCH] dialog context.

Log:

The content of a dialog: the inputs and outputs its participants exchange, as an ordered sequence of log entries per flow (Section 3.11). It corresponds to the log of the [ARCH] dialog context. Entry content is opaque to PACE; entry metadata is not.

Flow:

An ordered sequence of log entries that one participant sends in a dialog, with a single output class and delivery class, such as one participant's chat or audio. A binding defines how flows are named; in the MOQT binding, a flow is a track (Section 4.2.4).

Log position:

A value that orders the entries of one flow. Positions increase with each entry but need not be contiguous.

Dialog identifier (dialog-id):

A UUID that names a dialog for its whole lifetime.

Participant:

A user-facing client, agent, tool, or service that takes part in a dialog and processes its dialog context. Each participant has a participant identifier that is unique within the dialog. In the MOQT binding, the participant identifier is the agent-id field of the participant's tracks in the dialog namespace (Section 4.1). A participant plays the principal or agent role of [ARCH]; the controller is usually the principal's client.

Principal:

The user, workload, or organization on whose behalf a dialog runs and from whom an agent's authority derives, corresponding to the principal role of [ARCH]. A principal takes part through a participant, usually its client, or through an agent that acts for it; in a credential, the principal is the subject, not the agent presenting it (Section 7).

Invitee:

A participant in state invited: named in OPEN, or added by UPDATE, that has not yet answered with ACCEPT or REJECT (Section 3.3.1).

Controller:

The participant named by controller_id in the dialog context: initially the participant that opened the dialog, later an alternate controller that took over (Section 3.7.4). The controller is the authoritative source of the dialog context and the only participant that changes the participant set, expiration, generation, or terminal state.

Intermediary:

An entity that forwards dialog context between participants without being the source of it. A MOQT relay that forwards PACE tracks is an intermediary, and so is a gateway that transforms one modality into another.

Dialog host:

An intermediary that keeps a dialog's control messages and delivers them to its participants, in a binding where participants do not exchange them directly, such as a binding over HTTP. Unlike a relay, which forwards messages as they arrive, a dialog host stores them and serves them on request. A participant can also be the dialog host.

Propagation relationship:

The authorization and forwarding relationship that allows a participant or intermediary to receive or forward a dialog's context. PACE signals the creation and revocation of such relationships; the deployment's authorization mechanism enforces them.

Participant descriptor:

Optional members of a participant's entry in the dialog context that say what kind of party it is, whom it acts for, how it was authenticated and by whom, what it may receive, and what it has consented to (Section 3.4.1).

Posture:

Whether a participant treats the dialog as cooperative or adversarial (Section 3.4.3). Posture describes incentives, not hostility [DIALOG-UC].

Segment:

A span of a dialog whose content reaches only named participants, while the others learn that it happened and its result (Section 3.7.1).

MOQT session:

A transport-level session established with MOQT SETUP. A MOQT session is not a dialog: one dialog can span several MOQT sessions, and one MOQT session can carry several dialogs.

2. Architecture Overview

The PACE architecture layers as follows:

┌─────────────────────────────────────────────────────────┐
│  Application Protocols (A2A, MCP, AutoGen, Poppy, ...)  │
├─────────────────────────────────────────────────────────┤
│  PACE Dialog Management: dialog-id │ lifecycle │        │
│  propagation                                            │
├─────────────────────────────────────────────────────────┤
│  PACE MOQT Binding: namespaces │ control tracks │       │
│  media and operation tracks │ priority                  │
├─────────────────────────────────────────────────────────┤
│  MOQT: Pub/Sub │ Namespace Routing │ Relay Caching      │
├─────────────────────────────────────────────────────────┤
│  QUIC / WebTransport: TLS 1.3 │ Mux │ Migration         │
└─────────────────────────────────────────────────────────┘
Figure 1: PACE Architecture

Identity, authentication, authorization, and encryption are building blocks used across these layers. PACE integrates them (Section 7, Section 9) but does not define new ones.

2.1. Transport Requirements

Live dialogs (Section 1.1) place the following requirements on the transport that carries them. They refine the transport aspects of Section 8 of [ARCH] for live, multi-modal dialogs.

  • Low latency: Each flow reaches participants within the latency its modality needs: under 100 ms for real-time media, and under a second for chat and streaming model output (Section 5). A loss or a backlog on one flow does not add delay to another.

  • Push delivery: Content, new participants, new flows, and new dialogs reach participants as soon as they are published, without polling.

  • Interruption: Dialog control, such as an interruption or a cancellation, reaches its receiver ahead of content that is already queued. Real-time content that is too late to be useful is dropped instead of being delivered late.

  • Concurrent flows: Real-time media, chat, streaming model output, tool calls, and bulk transfers share one dialog and one connection, each with its own priority and reliability.

  • Fan-out: One message reaches every participant without the sender sending a copy to each. Participants that join late or reconnect can catch up.

  • Continuity: A dialog survives connection migration, network changes, reconnection, and intermediaries.

  • Scale and control: Dialogs scale through intermediaries across regions, and each intermediary enforces who may send and receive.

Request/response mechanisms might not meet these requirements. When a dialog's flows share one ordered connection, a cancellation waits behind output that has already been sent, and one lost packet stalls every flow. Delivery to several participants, and learning about new ones, are left to the application.

2.2. Why MOQT?

This document specifies a MOQT binding for exchanging dialogs. Other documents can bind the same dialogs to HTTP for agents that cannot use MOQT. Media over QUIC Transport (MOQT) [MoQ-TRANSPORT] is a publish/subscribe protocol over QUIC [RFC9000]. It meets each requirement in Section 2.1 with a transport mechanism, so agent protocols do not need to rebuild them.

  • Low latency: Objects are sent as soon as they are produced, on QUIC streams or datagrams alike, and relays forward them as they arrive, without waiting for the rest of a group. Each track has its own streams, so a loss on one flow does not hold up the others. Datagrams are not retransmitted, so real-time media keeps flowing under congestion instead of waiting for lost data.

  • Push delivery: Agents publish and subscribe directly, with no polling and no long-lived HTTP connections. Content reaches subscribers as soon as it is published. SUBSCRIBE_TRACKS on a namespace prefix delivers all current and future tracks under that prefix, so new participants and flows appear without a separate signaling channel. Output that arrives minutes or hours after its input reaches the subscriber without a webhook or a client-side HTTP server.

  • Interruption: Publisher priority (per subgroup or datagram) and subscriber priority (per subscription) let dialog control overtake queued lower-priority data, so a cancellation reaches the agent ahead of output already in flight. Per-track delivery timeouts let the publisher mark time-sensitive objects: objects that miss the deadline, such as a voice frame whose playout window has passed, are dropped at the relay instead of wasting bandwidth and queue depth.

  • Concurrent flows: One MOQT session carries sub-100ms media, sub-second chat, and best-effort bulk transfers at the same time, in one congestion control context. Objects can be delivered over QUIC datagrams, which suit real-time audio and video where lower latency matters more than retransmission, or over QUIC streams, which suit agent chat, operations, and bulk transfers that need ordered, reliable delivery.

  • Fan-out: A single PUBLISH reaches every subscriber through the relay, so multi-agent collaboration does not need application-level broadcast code, and the publisher keeps no per-subscriber state. Relay caches let late-joining or reconnecting agents catch up without the publisher retransmitting.

  • Continuity: QUIC provides connection migration (for example, from Wi-Fi to cellular) and 0-RTT resumption. A dialog is not tied to a MOQT session: one dialog can span several sessions (Section 1.3). An agent namespace is not tied to a host either, so the principal and the agent can each move between hosts over a dialog's lifetime and a new client instance catches up from relay caches.

  • Scale and control: Relays handle NAT traversal, load balancing, and fan-out without direct peer-to-peer connections, and they can be stacked hierarchically to spread load across regions without an all-to-all mesh. Namespace-level access control maps onto agent authorization scopes: the relay decides who may publish or subscribe to each namespace. QUIC provides TLS 1.3 for every hop.

Where UDP is blocked, the same MOQT binding can run over QMux [QMUX], which provides QUIC stream and datagram operations over TLS over TCP. Agents keep one protocol stack instead of falling back to a separate one, at the cost of TCP head-of-line blocking and reliable, ordered datagrams.

3. Agentic Dialog Management

This section defines the dialog context, its lifecycle, and the rules for propagating it. The mapping of this protocol onto MOQT is defined in Section 4.

3.1. Overview

A dialog has one controller, which holds the authoritative dialog context: the dialog's identifier, its state, its participants and their states, and a generation number. Other participants read that context and send requests; only the controller changes it.

Every change works the same way. A participant sends a request, such as ACCEPT to answer an invitation, LEAVE, or DECLARE to state what kind of party it is. The controller decides whether to apply it and, if it does, publishes UPDATE with the complete new context at the next generation. Receivers replace their copy of the context with each UPDATE instead of merging changes, and use the generation to detect a missed or stale update (Section 3.3.3).

The dialog is pending until an invitee accepts, then active, and finally terminated (Section 3.3). Each participant has its own state: invited, then active, and then left, rejected, or revoked (Section 3.3.1). A participant leaving does not by itself end the dialog.

A two-party dialog opens as follows:

  A (controller)                                         B
    |  OPEN    gen 0, pending, A active, B invited        |
    |---------------------------------------------------->|
    |  ACCEPT  in_reply_to: the OPEN                      |
    |<----------------------------------------------------|
    |  UPDATE  gen 1, active, A active, B active          |
    |---------------------------------------------------->|
    |                                                     |
    |      content: entries of the dialog log             |
    |<===================================================>|
Figure 2: Opening a Two-Party Dialog

The rest of this section defines, in order: the dialog context (Section 3.2); the dialog and participant states, the control operations, and how the controller commits changes (Section 3.3); what participants declare and how they are admitted and removed (Section 3.4); negotiated parameters (Section 3.5); stopping work and ending dialogs (Section 3.6); segments, delegated dialogs, and controller changes (Section 3.7); errors (Section 3.8); recovery after an interruption (Section 3.9); the encoding of control messages (Section 3.10); and the dialog log (Section 3.11).

3.2. Dialog Context

A dialog context contains the following fields:

  • dialog_id: A UUID version 4 [RFC9562]. The controller generates it when opening the dialog. It remains unchanged for the lifetime of the dialog, including when the dialog crosses an intermediary, changes modality, or resumes after a connection loss.

  • generation: An unsigned integer identifying the current controller state. The initial generation is 0. The controller increments it by one for each committed change to the dialog context, including a change in a participant's state.

  • state: One of pending, active, or terminated.

  • controller_id: The participant identifier of the controller.

  • participants: The set of participant identifiers in the dialog, each with a state: invited, active, left, rejected, or revoked.

  • expires_at: An optional absolute time in RFC 3339 date-time format. When present, participants treat the dialog as terminated at that time.

  • parent_dialog_ids: An optional list of dialog-ids, set by the controller in OPEN when the dialog is opened to carry out work delegated from another dialog. The first entry is the dialog that triggered this one, followed by that dialog's own parents, so the list traces the delegation chain back to the originating dialog (Section 6.2.1 of [ARCH]). It does not change after OPEN.

    A controller that opens a delegated dialog with participants in another trust domain MAY omit parent_dialog_ids, or list only the entries those participants may learn, so that parent dialog-ids are not disclosed across the boundary. It then keeps the link to the parent dialogs locally and applies lifecycle propagation (Section 3.7.2) and cancellation (Section 3.6.1) from that local link. Participants of such a dialog cannot detect a cycle in the dialog lineage; they rely on the delegation chain in credentials for cycle detection (Section 7.1.2).

  • related_dialogs: An optional list of objects, each with a dialog_id and a relation, set by the controller in OPEN and unchanged after it. relation is follow-up (the dialog carries updates on the outcome of the named dialog after it ended), resumes (the dialog continues a relationship established in the named dialog), or continues (the dialog carries on the same task on another channel). Unlike parent_dialog_ids, a related dialog is not delegated work: lifecycle propagation (Section 3.7.2) and cancellation do not follow this link, so a follow-up outlives the dialog it follows. A controller MAY withhold entries across a trust boundary, as for parent_dialog_ids.

  • alternate_controller_id: An optional participant identifier of an active participant that may take over as controller (Section 3.7.4).

The context is encoded as a JSON object:

{
  "dialog_id": "550e8400-e29b-41d4-a716-446655440000",
  "generation": 0,
  "controller_id": "agent-a",
  "state": "pending",
  "participants": [
    {"participant_id": "agent-a", "state": "active"},
    {"participant_id": "agent-b", "state": "invited"}
  ],
  "expires_at": "2026-10-23T12:00:00Z"
}

The expires_at member is omitted when the dialog has no configured expiration, parent_dialog_ids is omitted when the dialog was not opened from another dialog or when the controller withholds it (above), related_dialogs is omitted when there are none, and alternate_controller_id is omitted when no alternate is designated. A context snapshot MUST contain every other member, including participants whose states are left, rejected, or revoked. A controller MUST NOT remove a participant from a later snapshot; it changes that participant's state instead. The controller MUST appear in the participant set with state active.

The identity of the principal on whose behalf a delegated dialog runs travels in the credentials member as a delegation chain (Section 7.1.2), not in the dialog context.

These identifiers are correlation values. They MUST NOT be treated as credentials or bearer authorization. Knowing a parent dialog-id does not grant access to the parent dialog. A participant identifier MUST NOT encode personal or task information.

3.3. Lifecycle and Control Operations

The dialog lifecycle has three states:

                  OPEN (controller, generation 0)
                                |
                                v
                         +-------------+
              UPDATE +-->|   PENDING   |------------------+
                     |   +-------------+                  |
                     |      |      |                      |
                     +------+      | UPDATE committing    | TERMINATE
                                   | an invitee's ACCEPT  | or expiry
                                   v                      v
                         +-------------+            +------------+
      UPDATE, REVOKE +-->|   ACTIVE    |----------->| TERMINATED |
                     |   +-------------+ TERMINATE  |  (final)   |
                     |      |            or expiry  +------------+
                     +------+

Figure 3: Dialog State Machine
pending:

The controller has opened the dialog, but no invitee has accepted.

active:

At least one invitee has accepted, and the dialog has not been terminated or expired. The context can be modified with UPDATE and REVOKE.

terminated:

The controller has terminated the dialog, or its expiration time has passed. This state is final.

Only the controller causes transitions. UPDATE, REVOKE and TERMINATE each advance the generation by one. ACCEPT, REJECT, JOIN_REQUEST, LEAVE and DECLARE are requests the controller commits with UPDATE. Interruptions, cancellations, and segments do not change the state.

An interruption, such as a lost connection or relay failure, does not change the dialog state. Participants recover the current context as described in Section 3.9. A participant's own state is tracked separately from the dialog state: joining or leaving does not by itself terminate the dialog. When the controller commits a LEAVE that leaves it as the only active participant, it SHOULD publish TERMINATE. This lets an agent that is not the controller end a two-party dialog.

The control operations below cover the lifecycle functions of Section 6.1.2 of [ARCH]. Initiation is OPEN, and termination is TERMINATE, expiry, or a LEAVE that ends a two-party dialog. Suspension needs no operation: a dialog persists through idle periods with no transport connection, and a binding can pause individual flows. Resumption is STATE_REQUEST, which recovers the dialog context, followed by LOG_STATUS, which establishes what each receiver has accepted so that missed input and output are replayed without duplication (Section 3.11.3), and cancellation is CANCEL (Section 3.6.1).

PACE defines the following control operations:

Table 1: Control Operations
Operation Sender Generation Effect
OPEN Controller 0 Creates a pending dialog and identifies invitees.
ACCEPT Participant or invitee Current Accepts an invitation or a parameter proposal.
REJECT Participant or invitee Current Rejects an invitation or a parameter proposal.
UPDATE Controller Next Replaces the participant set, expiration, or controller.
JOIN_REQUEST Prospective participant Latest known, or 0 Requests admission to an existing dialog.
LEAVE Participant Current Leaves the dialog. The controller records the participant as left.
DECLARE Participant Current Declares the sender's own descriptors, or its observations of others (Section 3.4.2).
REVOKE Controller Next Revokes a participant's propagation relationship.
TERMINATE Controller Next Ends the dialog.
SEGMENT Controller Current Starts or ends a segment (Section 3.7.1).
CANCEL Any participant Current Asks receivers to stop work started by the sender's earlier input (Section 3.6.1).
CANCEL_ACK Any participant Current Reports the outcome of a CANCEL.
STATE_REQUEST Participant Current Requests the current context after a gap or reconnection.
STATE Controller Current Supplies a complete context snapshot.
LOG_STATUS Any participant Current Reports accepted log positions, or answers such a report (Section 3.11.3).
ERROR Any participant Current Reports that an operation was not applied, and why.

3.3.1. Participant States

Each entry in participants has its own state, separate from the dialog state. The controller sets it, in the same UPDATE, REVOKE, or OPEN that commits the change:

     named in OPEN, or added by UPDATE
     (an invitation or an admitted JOIN_REQUEST)
                      |
                      v
               +-------------+  REJECT committed   +------------+
               |   invited   |-------------------->|  rejected  |
               +-------------+                     +------------+
                      |
                      | ACCEPT committed
                      v
               +-------------+  LEAVE committed    +------------+
               |   active    |-------------------->|    left    |
               +-------------+                     +------------+
                      |
                      | REVOKE
                      v
               +-------------+
               |   revoked   |
               +-------------+
Figure 4: Participant State Machine

A participant in state left, rejected, or revoked stays in every later snapshot with that state (Section 3.2). Admission (Section 3.4.4, Section 3.4.5, Section 3.4.6) and revocation (Section 3.4.7) are defined in Section 3.4.

3.3.2. Committing Participant Changes

Only the controller changes the dialog context. Other participants send requests and responses, and the controller commits the result.

  • OPEN includes at least one invitee.

  • ACCEPT and REJECT answer a proposal from the controller, named by in_reply_to: an invitation in OPEN or UPDATE, or a parameter change carried in parameters on UPDATE. After the controller commits an invitee's answer, its state is active or rejected. A participant that rejects a parameter change remains in the dialog; the controller then withdraws the change in a later UPDATE or revokes the participant.

  • JOIN_REQUEST does not add a participant. The controller admits the requester by sending UPDATE with the requester's state set to invited, or refuses it with ERROR policy-rejected.

  • ACCEPT, REJECT, JOIN_REQUEST, and LEAVE do not modify the context by themselves. The controller commits each resulting change by publishing UPDATE with the complete context at the next generation.

  • The dialog becomes active when the controller publishes such an UPDATE after at least one invitee has accepted.

3.3.3. Generations and Ordering

UPDATE, REVOKE, and TERMINATE carry the next generation and a complete context, so a receiver replaces its state as a whole instead of merging partial updates. Other operations carry the current generation (Table 1).

  • The controller MUST serialize its state-changing operations and MUST NOT reuse a generation for different state.

  • A receiver MUST reject a controller operation whose generation is lower than the latest generation it accepted, with ERROR stale-generation.

  • If a receiver sees a generation greater than the next expected one, it MUST send STATE_REQUEST and MUST NOT infer the missing changes.

  • A prospective participant does not yet track the dialog, so its JOIN_REQUEST carries the latest generation it knows, or 0. The controller does not reject a JOIN_REQUEST as stale.

  • Receivers use the message identifier and generation to detect duplicates and stale messages. Reprocessing a duplicate operation MUST NOT repeat its side effects.

3.4. Participants

This section defines what participants declare about themselves and others, and how they are admitted and removed. Each change is a request that the controller commits with UPDATE (Section 3.3.2).

3.4.1. Participant Descriptors

In a dialog between a person and their own agent, the participants know who each other is. Many dialogs PACE carries are not like that [DIALOG-UC]. A consumer's agent negotiates with a brand's agent; a person on a phone call talks to a voice agent through a gateway; a recorder or a transcriber joins a meeting. Before a participant shares information or agrees to something, it needs answers to questions such as:

  • Is the other party a person or software?

  • For whom does it act, and what may it commit to?

  • How strongly was it authenticated, and by whom?

  • Has it agreed to be recorded, or to have its input processed?

  • Is it here to cooperate, or to bargain?

Policies, and in some places laws, on AI disclosure, recording consent, and delegated authority depend on these answers. Participant descriptors carry them in the dialog context, so that every participant sees the same answers and knows who gave each one.

Each entry in participants MAY carry the members below. They come from three sources: the participant itself, other participants, and the controller. The source matters, because a participant's statement about itself is not the same as another participant's observation of it.

Table 2: Participant Descriptors
Member Set by Value
kind The participant human, agent, tool, transformer, recorder, or gateway
acts_for The participant Opaque pseudonym of its principal
commit_scope The participant RAR authorization_details [RFC9396]: what it may commit to
consent The participant {"recording": s, "processing": s}; s is granted, refused, or withdrawn
posture The participant cooperative or adversarial (Section 3.4.3)
detected_kind Other participants List of {"kind", "by", "how"}; by is the observing participant, how is stated (the subject said so) or inferred
observed_consent Other participants List of {"recording", "processing", "by", "how"}, for a participant that cannot declare its own, such as a person on a call
assurance Other participants List of {"level", "by", "method", "at"}
access Controller {"send": [...], "receive": [...]}, naming modalities or flows
declaration_ref Controller message_id of the DECLARE whose values were committed
sponsor Controller Participant that sponsored this participant's admission (Section 3.4.5)
Declared members:

The controller copies them verbatim from the participant's latest DECLARE and names that DECLARE in declaration_ref. Any participant can check the copy against the DECLARE the sender published, which is authenticated by the sender's key-bound presentation (Section 7.1.4). A controller MUST NOT alter them. A participant that finds a mismatch MUST disregard the controller's copy.

Observed members:

Each entry names the participant that made it, in by, so that every participant decides for itself whether to rely on it. This is what separates a participant's claim to be a person (kind) from another participant's detection of automation (detected_kind).

assurance:

How strongly a participant was authenticated. level is none (a guest, not authenticated), asserted (identity claimed but not verified), authenticated, or delegated (authenticated, with verified authority from a principal). method is a value from the PACE Assurance Methods registry (Section 12): credential, a key-bound presentation (Section 7.1.4) of a credential mechanism (Section 7.1.1); cimd, a key-bound presentation checked against a key from an OAuth Client ID Metadata Document [CIMD]; resumed, keys pinned in an earlier dialog named in related_dialogs with relation resumes; or join-code (Section 3.4.6). Other documents can register further methods, such as those for telephony.

access:

What the participant may send and receive, such as audio but not screen share. It differs from the supported_modalities parameter, which states what a participant can handle. Absent access, a participant may send and receive every flow. Relays and dialog hosts enforce it (Section 9.2.2).

An example entry, for a consumer's agent as seen in a dialog with a brand's agent:

{"participant_id": "p-7c1e", "state": "active",
 "kind": "agent", "acts_for": "u-4f9a",
 "detected_kind": [{"kind": "agent", "by": "brand-agent"}],
 "assurance": [{"level": "delegated", "by": "brand-agent",
                "method": "credential",
                "at": "2026-11-13T10:02:11Z"}],
 "posture": "adversarial",
 "declaration_ref": "9b2f6c1e-0d4a-4e8b-a7c3-5f1d2e3a4b6c"}

3.4.2. Declarations

A participant states its own descriptors, and what it observed about others, with DECLARE. DECLARE is a separate operation, rather than a field the controller fills in, because the controller may act for the other side. A brand's agent that controls a dialog should not be able to say that the consumer's agent consented to recording, or that the consumer's agent is a person. With DECLARE, the statement comes from the participant and is signed by it; the controller only copies it, and anyone can check the copy.

The declaration member of DECLARE is an object with any of these members:

Own descriptors:

kind, acts_for, commit_scope, consent, and posture. Each replaces the sender's previously declared value.

sponsored:

A list of participant identifiers whose admission the sender sponsors (Section 3.4.5).

observations:

A list of objects, each with subject (a participant identifier), one or more of detected_kind, assurance, and consent, and optionally how (stated or inferred). Each states what the sender concluded about that participant, or what the participant said when it cannot declare for itself, such as a person on a call who agrees to be recorded.

disclosure:

A record that the sender disclosed that it is an AI, with at (the time), to (the participant identifiers it was made to), and entry (the entry reference of the log entry that carried it). It records what the sender said, not that it was understood.

The controller commits a DECLARE with UPDATE at the next generation. In that UPDATE, it:

  1. copies the sender's own values into the sender's entry and sets declaration_ref; and

  2. appends each observation to the subject's detected_kind, assurance, or observed_consent, with by set to the sender. An observed consent never replaces a participant's own consent declaration.

The controller MUST commit a well-formed DECLARE from an active participant without undue delay. It MUST NOT refuse a declaration of kind, a raise of posture, or a consent value of refused or withdrawn. A participant must be able to say what it is, to become more cautious, and to refuse recording, without the agreement of a controller that may act for its adversary. A controller MAY refuse other declarations with ERROR policy-rejected.

3.4.3. Posture

Many agent dialogs are between parties whose interests differ: a buyer's agent and a seller's agent, or a customer's agent and a support agent handling a dispute. In such a dialog, an agent should not share everything it knows about its principal, and should not rely on the other side's statements without checking them. Posture lets a participant say which kind of dialog it considers this to be, so that every participant applies the matching policy and knows that the others have been told. Posture describes incentives, not hostility.

A participant declares its posture with DECLARE:

  • The dialog's effective posture is adversarial if any active participant declares it, and cooperative otherwise.

  • Each participant decides from its own policy what posture means for it, such as whether a routine exchange may proceed without a per-action approval (Section 9.1.2).

  • A counterpart's declaration of cooperative MUST NOT lower the assurance, approvals, or records that a participant's policy requires.

A participant can raise its posture at any point, for example when an order-status inquiry turns into a refund dispute. The controller commits the change like any declaration and cannot refuse it (Section 3.4.2), so every participant learns of it. Context shared before the change stays shared; after it, participants SHOULD stop sharing context they would not give an adversary, and each applies its adversarial policy to later exchanges.

3.4.4. Guests and Assurance Upgrade

A controller MAY admit a participant that has not authenticated, such as a caller on a channel that carries no credential, and records its assurance level as none. The deployment's policy limits what a guest may do, such as receiving general information only. The participant raises its level later in the same dialog by presenting credentials on a later control message, such as a DECLARE; the controller, or another participant that verifies them, records a new assurance entry. No new dialog is needed. A participant that requires a higher level for an operation rejects it with ERROR insufficient-assurance, stating what it needs (Section 3.8.1); a participant that cannot or will not upgrade answers with ERROR upgrade-declined, and the dialog continues at the current level.

3.4.6. Join Codes

A participant on a channel that cannot carry the dialog identifier, such as a voice call over the PSTN, can bind a new channel to the same dialog with a join code: a short-lived, single-use value that the controller issues to an existing participant and that is conveyed in the call. A JOIN_REQUEST MAY carry join_code. If it is valid, the controller admits the requester as the sponsored participant of the participant to which the code was issued, records an assurance entry with method join-code, and invalidates the code. A code MUST have at least the entropy its lifetime and the controller's rate limit on attempts require, and a controller MUST limit attempts per dialog. A requester that does not know the dialog-id redeems the code at an endpoint that resolves it, such as one that a binding over HTTP provides. Other documents can define encodings for voice and DTMF.

3.4.7. Revoking a Participant

REVOKE is controller-authoritative. It carries the complete context at the next generation, sets the target participant's state to revoked, and ends that participant's propagation relationship. A participant or intermediary that receives REVOKE:

  • MUST stop forwarding the dialog's context to the target.

  • MUST cancel subscriptions that it holds only to propagate the dialog to the target.

The revoked participant MUST stop acting under the authorization it held in the dialog. It MUST send TERMINATE in each delegated dialog it controls whose parent_dialog_ids include the dialog, so that work it delegated under that authorization stops as well (Section 3.7.2). Revocation of the participant's credential itself is a function of the credential mechanism (Section 7.1.1).

3.5. Dialog Parameters

Dialog parameters are negotiated in the parameters member of OPEN, ACCEPT, UPDATE, and JOIN_REQUEST (Section 3.10). They describe how the dialog is conducted and are not part of the dialog context:

  • dialog_type: Short-lived (bounded duration) or long-lived (indefinite)

  • idle_timeout: Maximum idle duration before the controller terminates the dialog

  • max_participants: Maximum number of participants in the dialog

  • supported_modalities: Data modalities the participant can handle

  • supported_protocols: Application protocols offered (A2A, MCP, etc.)

  • content_encoding: Compression applied to content payloads (for example, gzip); absent means none (Section 4.2.3)

  • liveness_interval: Maximum interval, in seconds, between liveness signals from a participant doing work in the dialog (Section 3.7.3)

  • controller_capable: Whether the participant can act as controller if designated as alternate (Section 3.7.4)

  • log_retention: Minimum time, in seconds, for which a participant retains entries of its reliable flows that a consumer has not yet accepted (Section 3.11.3)

  • purpose: On OPEN and JOIN_REQUEST, a short statement of why the sender opens or joins the dialog, such as a delivery update or a notetaker for a meeting

  • basis: On OPEN and JOIN_REQUEST, why the sender considers itself entitled to make contact, from the PACE Contact Basis registry (Section 3.5.1)

Parameters are negotiated by proposal and acceptance:

  • The controller proposes values in OPEN, or in UPDATE for a change.

  • Each invitee or participant accepts with ACCEPT, stating the values it supports. A participant that cannot accept a proposal sends REJECT.

  • For list-valued parameters, such as supported_modalities and supported_protocols, the accepted values MUST be a subset of the proposed ones, and they are the participant's active set. Accepting no value of supported_modalities is a rejection.

  • For scalar parameters, the accepted value MAY lower a limit (max_participants, idle_timeout, liveness_interval) but MUST NOT raise it. log_retention is a minimum, so the accepted value MAY raise it but MUST NOT lower it.

  • purpose and basis are not negotiated.

  • The controller's committing UPDATE carries the resulting values in parameters, and those values are in effect from that generation.

3.5.1. Contact Purpose and Basis

An agent can open a dialog with a party that did not ask for it, such as a person's agent or an organization's inbox. The recipient needs to decide whether to engage before any content is exchanged, as a person screens a call or an email [DIALOG-UC]. purpose says what the contact is about, and basis says why the sender considers itself entitled to make it. With contact_limits (Section 3.6.3), they let a recipient refuse a contact, say why, and limit later ones.

basis takes one of these values:

subscription:

The recipient asked for contacts of this kind, such as updates on an order or a delivery.

existing-relationship:

The sender and the recipient's principal have an ongoing relationship, such as a customer account, and the contact concerns it.

contract:

The contact is needed to carry out an agreement to which the recipient's principal is a party, such as arranging the delivery of something it bought.

legal-obligation:

Law or regulation requires the sender to make the contact, such as a product safety recall.

consent:

The recipient's principal agreed to contacts of this kind from this sender, outside any subscription.

none:

The sender claims no basis: the contact is unsolicited, and says so.

purpose and basis are the sender's assertions, and the recipient checks them. A recipient can check subscription, and often existing-relationship and consent, against its own records. It cannot in general check contract or legal-obligation; it records them and weighs them with the sender's assurance (Section 3.4.1). A recipient can refuse a contact whose basis it cannot confirm, or that states no purpose or basis, with REJECT and a result reason such as unwanted-contact or no-stated-purpose (Section 3.6.3). The values describe the sender's reason for making contact; they are not claims of a legal basis for processing personal data under any jurisdiction.

3.5.2. Dialog Lifetime

Short-lived dialogs carry an expires_at in the dialog context and end when it passes. Long-lived dialogs omit expires_at and continue across idle periods and interruptions until the controller terminates them. When idle_timeout is set and the dialog has been idle for that long, with no content or control sent by any participant, the controller SHOULD publish TERMINATE.

3.6. Stopping Work and Ending Dialogs

CANCEL stops work without ending the dialog (Section 3.6.1); TERMINATE and expiry end it (Section 3.6.2); and results say how a participant regards the outcome (Section 3.6.3).

3.6.1. Cancellation

A principal that interrupts an agent, a policy decision, or a failed subtask can require work to stop at once. PACE carries cancellation as a dialog control operation rather than relying on connection teardown, so a binding can deliver it ahead of queued content. The MOQT binding sends it on the ctrl track at critical priority (Section 4.3.6).

CANCEL does not change the dialog context and does not end the dialog. Its scope member selects the work to stop:

operation:

A single operation the sender started, named in target.

subtree:

All work the receivers carry out in delegated dialogs whose parent_dialog_ids include this dialog. A receiver that controls such a dialog sends CANCEL with scope dialog in it, so cancellation propagates down the delegation chain and dependent subtasks do not continue after the parent work is withdrawn.

dialog:

All work the sender started in this dialog, together with its subtree.

A receiver applies CANCEL through the application protocol's own cancellation where one exists, such as MCP notifications/cancelled or A2A CancelTask, so the application sees its native cancel. CANCEL is idempotent: a receiver that has already acted on a CANCEL reports the same outcome again and does not repeat side effects.

Each receiver doing affected work answers with CANCEL_ACK, giving one of these outcomes:

  • stopped: the work stopped before completing.

  • completed: the work had already completed.

  • partially-reversed: the work stopped, and some but not all of the actions it took were rolled back.

  • irreversible: the work stopped, but an action it took could not be rolled back.

  • unknown: the receiver cannot determine whether the work, or work it delegated, stopped.

CANCEL_ACK MAY also carry compensation, an object with a description and an opaque reference, for what the receiver offers in place of an action it could not reverse, such as a refund or a credit. The terms under which work can be cancelled before it commits are stated in a commitment (Section 9.1.2).

When CANCEL reaches delegated dialogs (scope subtree or dialog), a receiver that controls such dialogs sends its own CANCEL_ACK only after it has received a CANCEL_ACK from each participant doing work in them, or after a locally configured timeout. Its outcome is the least favorable among its own and those it received, in the order completed, stopped, partially-reversed, irreversible, unknown; a delegated participant that did not answer before the timeout counts as unknown. The initiator of a CANCEL therefore learns whether the cancellation took effect along the whole chain, or that this could not be determined.

Such a receiver MAY also report each delegate's outcome in details, a list of objects with delegate, outcome, and optionally dialog_id and compensation. delegate is the delegate's participant identifier, or an opaque label where the receiver withheld the delegated dialog from the initiator (Section 3.2); in that case dialog_id is omitted. A delegate that did not answer is listed with unknown.

3.6.2. Termination and Expiry

TERMINATE is final.

  • After TERMINATE, a receiver MUST reject later operations for the dialog with ERROR dialog-terminated, except a duplicate TERMINATE or a STATE needed to complete synchronization.

  • Expiration has the same effect as TERMINATE. Each participant evaluates it locally from expires_at, and it does not create a new generation.

  • A controller SHOULD publish TERMINATE before the expiration time when it can.

  • Resuming a transport connection does not reopen a terminated or expired dialog.

3.6.3. Results and Contact Limits

ERROR reports that an operation failed. A dialog that ends normally can still end without what a participant wanted: a refund declined, an agreement not reached, a contact refused. TERMINATE, LEAVE, and REJECT MAY carry result, which states how the sender regards the dialog's outcome:

"result": {
  "category": "deferred",
  "reason": "policy",
  "next_update_at": "2026-11-20T09:00:00Z",
  "recourse": "https://brand.example/disputes",
  "parts": [{"label": "pickup", "category": "completed",
             "compensable": false},
            {"label": "refund", "category": "deferred",
             "compensable": true}],
  "sources": ["a1b2c3d4-..."],
  "commitments": ["6f1c0e9a-..."]
}
  • category: REQUIRED. One of completed, declined, deferred, partial, transferred, failed, or abandoned.

  • reason: OPTIONAL. A value from the PACE Result Reasons registry (Section 12), such as ineligible, suspected-fraud, or no-stated-purpose.

  • retry_after: OPTIONAL. The time before which the recipient should not ask again.

  • recourse: OPTIONAL. A URI at which the recipient can pursue the matter, such as an appeal or a human channel.

  • next_update_at: OPTIONAL. For deferred, when the sender expects to report next, so that a waiting participant does not poll.

  • parts: OPTIONAL. For partial, the outcome of each part, and whether a part that took effect can be compensated.

  • sources and commitments: OPTIONAL. Identifiers of attested assertions and commitments (Section 9.1.2) on which the result rests, so that each fact keeps its source.

result is the sender's statement. It does not change the dialog state beyond the operation it accompanies, and a participant that needs to rely on it checks the assertions and commitments it cites.

REJECT, LEAVE, and TERMINATE MAY also carry contact_limits, which asks the recipient to limit further contact:

  • retry_after: The time before which the recipient does not contact the sender again.

  • max_contacts and period: At most max_contacts contacts in any period seconds.

  • scope: dialog, for dialogs related to this one, or sender, for any dialog from the same sender.

  • opt_out: When true, the sender refuses further contact of this kind altogether; later contacts are refused with ERROR unwanted-contact.

Limits apply to the recipient as identified by its credentials. They cannot be tied to a participant that has not authenticated (Section 3.4.4). They are enforced by the sender's own policy; a recipient that ignores them is refused.

3.7. Segments, Delegated Dialogs, and Controller Changes

These rules apply to dialogs whose content some participants may not see (Section 3.7.1), dialogs opened to carry out work delegated from another dialog (Section 3.7.2, Section 3.7.3), and dialogs whose controller changes (Section 3.7.4).

3.7.1. Segments

Part of a dialog sometimes has to reach only some participants: card details entered for a payment tool, an off-the-record span in a meeting, or a challenge that only the person behind an agent may answer [DIALOG-UC]. A dialog has one visibility scope (Section 6), so the content of such a segment is exchanged in a separate dialog whose parent_dialog_ids names this one, and whose participants are only those who may see it. This dialog, the parent, carries only the segment's start, end, and result.

The controller announces a segment with SEGMENT. Its segment member has:

  • segment_id: A UUID version 4.

  • kind: masked, for content delivered only to the named participants, such as payment data, or off-the-record, for a span that agents and recorders do not receive.

  • phase: start or end.

  • participants: The participants that receive the segment's content.

  • dialog_id: OPTIONAL. The dialog that carries the content. The controller MAY withhold it from participants not listed.

  • result and reference: OPTIONAL, on end. The segment's result (Section 3.6.3) and an opaque reference to it, such as a payment reference, that does not expose its content.

Each participant not listed in participants MUST acknowledge a SEGMENT with ACCEPT naming it in in_reply_to. A recorder records the gap, so its record shows that a segment happened, its kind, and its result. For off-the-record, the controller MUST NOT admit to the segment dialog any participant whose kind is agent or recorder, and the segment's participants exchange its content only in the segment dialog. Until a participant has acknowledged, it might not know that a gap is intended; the controller MAY revoke a participant that does not acknowledge, and a dialog host withholds later content from it.

Because the segment dialog is a delegated dialog of this one, it ends when this one ends (Section 3.7.2). With end-to-end protection (Section 9.2.1), it has its own keys, so participants of this dialog that are not in the segment cannot read its content even through the relays they share.

3.7.2. Lifecycle Propagation Along Delegation Chains

A delegated dialog carries out work for the dialogs listed in its parent_dialog_ids. Its lifecycle follows theirs:

  • When a dialog terminates or expires, each participant that controls a delegated dialog whose parent_dialog_ids include it MUST send TERMINATE in that delegated dialog. Termination therefore reaches the end of the delegation chain, as CANCEL with scope subtree does (Section 3.6.1).

  • When a participant is revoked, it terminates the delegated dialogs it controls under that participation (Section 3.4.7).

  • A controller that terminates a delegated dialog for one of these reasons and needs to know whether work below it stopped first sends CANCEL with scope dialog and waits for the CANCEL_ACK outcomes (Section 3.6.1).

These rules use only the linkage in parent_dialog_ids, or the controller's local link when it withheld that member (Section 3.2). A participant that cannot reach a delegated dialog's other participants still terminates it; they observe the termination when they resynchronize.

3.7.3. Liveness

When the liveness_interval parameter is set, a participant that is doing work in the dialog for another participant publishes a liveness signal at least once per interval, even when it has no other output. In the MOQT binding, the signal is an object on its progress track (Section 4.1.2); any other object the participant publishes in the dialog also counts.

If a delegating participant receives no liveness signal from a delegated participant for two consecutive intervals, it treats that participant as unavailable: it halts the affected work, sends CANCEL for it, and MAY ask the controller to revoke the participant. Unavailability is not revocation; it can be a transient failure, and the participant can resume (Section 3.9).

3.7.4. Controller Transfer

A dialog's controller can change in two ways. In both, the new controller is the participant named in alternate_controller_id, which the controller sets with UPDATE and which MUST name an active participant that set controller_capable in its ACCEPT parameters.

  • Hand-off: The controller publishes UPDATE at the next generation with controller_id set to the alternate. From that generation, the alternate is the controller and the former controller is an ordinary participant.

  • Takeover: If the controller is unreachable (Section 3.9) for longer than the deployment's takeover time, the alternate publishes UPDATE at the next generation after the last one it accepted, with itself as controller_id, and with credentials showing that it is authorized to control the dialog, such as a delegation from the principal that authorized the original controller.

Participants accept a takeover UPDATE only from the participant named in alternate_controller_id in the context they hold, and only after verifying its credentials. If a participant receives two different contexts for the same generation, one from the former controller and one a takeover by the alternate, it MUST accept the takeover. A participant that has already accepted a later generation from the former controller rejects the takeover with ERROR stale-generation; an alternate that receives this withdraws, resynchronizes with STATE_REQUEST, and does not act as controller. A former controller that sees an accepted context naming another controller MUST stop publishing state-changing operations and resynchronize with STATE_REQUEST.

Controller transfer keeps the dialog-id and generation sequence, so a dialog continues through the failure of a controller, including an application-level mediator that controls it, rather than being replaced.

Editor's note: Takeover by an alternate trades availability for the risk of a brief split between controllers. Deployments that cannot tolerate this omit alternate_controller_id.

3.8. Authorization and Errors

Before applying an operation, a participant MUST authenticate its sender and check authorization using the deployment's mechanisms (Section 9). A valid state-changing operation needs no separate acknowledgment. An ERROR names the rejected control message in in_reply_to, or the rejected application operation in target (Section 3.8.1), and gives a reason.

  • A message that cannot be parsed, or that lacks a required member (Section 3.10), is rejected with ERROR malformed.

  • If the sender's credential is missing or fails verification, the participant MUST reject the operation with unauthenticated, or with credential-invalid when the credential is revoked or expired.

  • If authorization fails or cannot be determined, the participant MUST reject the operation with ERROR, giving unauthorized or authorization-unavailable as the reason, or scope-exceeded when the operation exceeds the scope delegated to the sender.

  • A participant that does not understand the control-message version or an operation MUST NOT reinterpret it. It responds with ERROR, giving unsupported-version or unsupported-operation.

Reasons are precise enough for the sender to decide what to do next, but an ERROR MUST NOT reveal policy details beyond the reason.

3.8.1. Error Reasons

The reason member of ERROR takes one of the values below. They distinguish validation, authentication, authorization, policy, state, and processing failures from transport failures, which MOQT reports separately (Section 6.1.5 of [ARCH]).

Table 3: ERROR Reasons
Reason Meaning
malformed The message cannot be parsed or lacks a required member.
unsupported-version The control-message version is not supported.
unsupported-operation The operation type is not supported.
stale-generation The generation is older than the latest accepted.
unauthenticated The sender's credential is missing or fails verification.
credential-invalid The sender's credential is revoked, expired, or otherwise invalid.
unauthorized The sender is not authorized for the operation.
authorization-unavailable Authorization could not be determined at this time.
scope-exceeded The operation exceeds or contradicts the scope delegated to the sender.
policy-rejected Local policy refuses the operation, such as a join.
delegation-cycle The delegation chain or dialog lineage contains a cycle (Section 7.1.2).
dialog-terminated The dialog is terminated or expired.
participant-revoked The sender has been revoked from the dialog.
validation-failed The operation was refused because it could cause unintended or irreversible effects.
insufficient-assurance The sender's assurance is below what the operation needs; ERROR carries required_assurance (Section 3.4.4).
upgrade-declined The sender cannot or will not raise its assurance; the dialog continues at the current level.
unwanted-contact The recipient refuses contact from the sender (Section 3.6.3).
translation-failed A mediator could not translate the operation or its result between protocols.
timeout The operation, or a step it depends on, did not complete in time.
internal-error The receiver failed while processing the operation.

The first group of reasons applies to control messages. A participant that rejects an application operation, such as a mediator that refuses or cannot translate a tool call, also MAY publish ERROR with target naming the operation, in the form CANCEL uses (Section 4.3.6), and SHOULD also answer in the application protocol where that protocol has a matching error.

Errors that the application protocol reports for its own operations, such as a failed tool call (an MCP error result) or a failed task (an A2A task in the failed state), stay in that protocol. A participant can therefore tell a transport failure (MOQT), a dialog failure (PACE ERROR), and a tool or agent processing failure (application protocol) apart by where each is reported.

3.9. Propagation and Continuity

  • Correlation across hops: An intermediary MUST preserve the dialog identifier, generation, and lifecycle operation when forwarding dialog context. It MUST NOT create a new dialog identifier or change the dialog state. This document defines no identifier-alias mechanism; a deployment that needs aliases requires a separate mechanism that securely correlates the alias with the original dialog.

  • Trust boundaries: A participant forwarding dialog context into another trust domain does so only within a propagation relationship that the deployment's authorization mechanism permits. Authorization to forward or subscribe at the transport layer does not by itself make an entity a participant.

  • Modality changes: Transforming a modality, such as converting audio to text, does not create a new dialog. Participants and intermediaries MUST keep the same dialog identifier and current lifecycle state across the transformation.

  • Continuity over time: After a connection loss, a participant MAY reconnect, over a new transport session if necessary, using the existing dialog identifier. It MUST obtain the current context from the controller with STATE_REQUEST before acting on state that may have changed while it was disconnected, and it uses LOG_STATUS to establish which log entries were accepted (Section 3.11.3). Acceptance of an entry does not imply that the application operation it carries completed; entries a receiver has not accepted are replayed, and entries it has accepted are not.

  • Unreachability: If the controller is unreachable, participants MUST suspend actions that depend on an unconfirmed context change. Each participant remains at the last generation it accepted until the controller is reachable again, an alternate controller takes over (Section 3.7.4), or the dialog expires. Participants MUST NOT create a replacement dialog solely because a path or intermediary failed. If an invitee or participant is unreachable, the controller MAY revoke it or let the dialog expire, according to deployment policy.

3.10. Control Message Encoding

A PACE control message is a UTF-8 encoded JSON object [RFC8259]. Every control message contains these members:

  • version: The integer 1.

  • type: One of the operations in Section 3.3.

  • dialog_id: The canonical textual form of the dialog's UUID.

  • generation: A non-negative integer no greater than 2^53 - 1.

  • sender_id: The sender's participant identifier.

  • message_id: A UUID version 4 unique to this message.

Control messages also carry the following type-specific members:

  • context: The complete dialog context, encoded as the JSON object defined in Section 3.2.

  • in_reply_to: The message_id of the control message this message answers: the OPEN, UPDATE, or SEGMENT that an ACCEPT or REJECT answers; the STATE_REQUEST or LOG_STATUS that a STATE or LOG_STATUS answers; the CANCEL that a CANCEL_ACK answers; or the message that an ERROR rejects.

  • target_id: The participant identifier whose propagation relationship is revoked.

  • declaration: The sender's declaration (Section 3.4.2).

  • segment: The segment being started or ended (Section 3.7.1).

  • scope: The cancellation scope: operation, subtree, or dialog (Section 3.6.1).

  • target: The operation to cancel when scope is operation, or the application operation an ERROR rejects. Its form is defined by the transport binding (Section 4.3.6).

  • outcome: One of stopped, completed, partially-reversed, irreversible, or unknown.

  • last_accepted_generation: The highest generation the sender has accepted from the controller, or -1 if it has accepted none.

  • reason: One of the values in Table 3.

  • positions: A list of objects, each with sender_id, flow, and accepted, giving the accepted position for one flow (Section 3.11.2). The forms of flow and accepted are defined by the binding.

Each type requires the members listed below:

Table 4
Type Required type-specific members
OPEN context
ACCEPT in_reply_to
REJECT in_reply_to; reason optional
UPDATE context
JOIN_REQUEST none
LEAVE none
DECLARE declaration
REVOKE target_id, context
TERMINATE context
SEGMENT segment
STATE_REQUEST last_accepted_generation
STATE context; in_reply_to when answering a STATE_REQUEST
LOG_STATUS positions; in_reply_to when answering a LOG_STATUS
CANCEL scope; target when scope is operation
CANCEL_ACK in_reply_to, outcome
ERROR reason; in_reply_to or target

For OPEN, context has generation 0 and state pending, and carries parent_dialog_ids when the dialog is delegated. For UPDATE, REVOKE, and TERMINATE, context has the next generation, and a TERMINATE context has state terminated. An UPDATE that adds a participant sets that participant's state to invited; the participant becomes active only after it accepts and the controller publishes a later UPDATE.

Further optional members are defined for some types:

  • result and contact_limits: On TERMINATE, LEAVE, and REJECT (Section 3.6.3).

  • required_assurance: On ERROR with reason insufficient-assurance, an object with level (Section 3.4.1) and optionally acr_values and max_age with their OpenID Connect meanings, for the authentication of the sender's principal.

  • sponsor and join_code: On JOIN_REQUEST (Section 3.4.5, Section 3.4.6).

  • compensation and details: On CANCEL_ACK (Section 3.6.1).

Three optional members carry information that is not part of the dialog context:

  • parameters: On OPEN, ACCEPT, UPDATE, and JOIN_REQUEST, the dialog parameters being proposed or accepted, such as modalities and application protocols (Section 3.5).

  • credentials: On any message, an identity assertion or authorization proof in a format selected by the deployment (Section 7.1.1). A participant refreshes its credentials before they expire by including new ones on a later message.

  • nonce: On OPEN, ACCEPT, UPDATE, JOIN_REQUEST, STATE_REQUEST, and STATE, a fresh random value, with at least 128 bits of entropy and base64url-encoded, to which the recipient binds the presentation in its next credentials (Section 7.1.4).

JSON member order is insignificant. Receivers MUST ignore unrecognized optional members, MUST reject an unrecognized type, and MUST reject messages with missing required members, invalid member types, or an unsupported version. Receivers MUST limit the size of a control message to a locally configured maximum. A participant MUST NOT include application content in a control message.

3.11. Dialog Log

The dialog context tells participants which dialog they are in and what state it is in. The log tells them what has been said in it. PACE does not interpret log content, but it defines the metadata of each log entry, so that participants can agree on what was exchanged when they resume, recover, or start a new client instance (Section 4.2.10 of [ARCH]).

3.11.1. Log Entries

Each log entry has the following metadata:

  • sender_id: The participant identifier of the sender.

  • flow: The flow that carries the entry. Its form is defined by the binding.

  • position: The entry's log position in that flow. Its form is defined by the binding. A sender MUST assign positions that increase with each entry it sends on a flow and MUST NOT reuse a position for different content.

  • timestamp: Optionally, the time at which the sender produced the entry. Receivers MUST NOT rely on timestamps from different senders being comparable.

  • source_id: Optionally, for an entry that renders another participant's input, such as a translation or a transcription, the participant identifier of the participant whose input it renders.

  • derived_from: Optionally, the entry references of the entries it was derived from.

sender_id, flow, and position together form an entry reference, which names one entry for the lifetime of the dialog.

A transformer, such as an interpreter or a speech-to-text gateway, is a participant (Section 3.4.1). It publishes each rendering on a flow of its own, so that different participants receive different renderings of the same exchange by choosing which flow to consume, and records are kept with the original apart from its renderings. Its renderings carry source_id and derived_from, so they are attributed to the speaker as derived content and not mistaken for the speaker's own words. Content a transformer sends that renders no input, such as a note of its own, carries neither member and is attributed to the transformer. PACE cannot check that a rendering is faithful; it keeps each rendering linked to its source.

Each flow has an output class and a delivery class, fixed for the life of the flow and defined by the binding's flow naming:

  • Output class (Section 4.2.11 of [ARCH]): input, for content a participant sends in the principal role; response; information, such as tool calls, sub-agent invocations, and reasoning summaries; or progress, such as token counts, elapsed time, and liveness signals.

  • Delivery class: reliable, for content that every receiver needs, such as chat, operations, and their results; or loss-tolerant, for real-time media whose value expires, such as audio and video frames.

Entries in a reliable flow are delivered in position order. Entries in a loss-tolerant flow can be lost, and later entries do not depend on earlier ones.

3.11.2. Acceptance

A receiver accepts an entry when it takes responsibility for it: from then on it either processes the entry or reports an ERROR for it. The accepted position of a flow, for one receiver, is the highest position in that flow up to which the receiver has accepted every entry it needs to.

  • A receiver MUST record an entry as accepted before it, or the agent it serves, begins to act on the entry, and MUST keep that record for as long as it can resume the dialog. A receiver that fails after acting on an entry therefore never reports it as unaccepted.

  • A receiver MUST discard an entry at or below its accepted position for that flow as a duplicate, without repeating its side effects.

  • For a loss-tolerant flow, the accepted position is the highest position received; earlier lost entries are not recovered.

Acceptance is not completion. An accepted request can still be in progress, fail, or be cancelled (Section 3.6.1); the application protocol reports those outcomes.

3.11.3. Log Status and Replay

After a gap, reconnection, or failover, participants exchange their accepted positions with LOG_STATUS so that each sender knows what to make available again, and each receiver knows what to retrieve. This answers whether a sender needs to replay its input (Section 6.1.2 of [ARCH]) without inferring it from transport state.

  • A participant that resumes a dialog first resynchronizes the dialog context with STATE_REQUEST (Section 3.9). It then sends LOG_STATUS without in_reply_to, listing in positions its accepted position for each flow it consumes.

  • Each participant that receives such a LOG_STATUS and consumes flows of the sender answers with LOG_STATUS, with in_reply_to naming it, listing its accepted positions for the sender's flows only.

  • For each reliable flow it sends, the resuming participant makes every entry above the lowest accepted position reported for that flow available to the receivers again, at the same positions and with the same content, as the binding defines. It MUST NOT re-send an entry with new content at an existing position, and MUST NOT send again as new entries the content of entries a receiver has already accepted.

  • The resuming participant retrieves entries above its own accepted positions in other participants' flows, as the binding defines.

  • Entries in loss-tolerant flows are not replayed.

A sender retains each entry of a reliable flow until every participant consuming that flow reports an accepted position at or above it, or for the log_retention time (Section 3.5), whichever comes first. A receiver whose accepted position falls below what the sender still retains cannot recover the missing entries; it SHOULD report this to the application and MAY request the current state through the application protocol.

A new client instance of a participant (Section 4.3.5.1) uses the same procedure to rehydrate the log: it retrieves the entries it needs to show, in each flow, from the positions it chooses. A binding MUST define how a participant retrieves entries by flow and position range, or MUST state that it does not support rehydration.

LOG_STATUS does not change the dialog context and does not advance the generation. A participant MUST NOT include log content in it.

4. MOQT Binding

This section specifies how the dialog management protocol of Section 3 is carried over MOQT.

4.1. Namespace Organization

PACE uses MOQT's hierarchical namespace for routing, isolation, and multiplexing. In MOQT, content is addressed by a Track Namespace (an ordered set of up to 32 fields, compared as byte sequences [MoQ-TRANSPORT]) together with a Track Name. PACE organizes communication into two namespace scopes: an agent namespace for each agent's identity and invitations, and a dialog namespace for all of a dialog's traffic.

4.1.1. Agent Namespace

  ("pace", "agents", {agent-id}) -- {track}

This is an agent's persistent identity on the relay network. Each agent publishes the following tracks:

  • "card" — Agent Card containing capabilities, identity claims, authentication schemes, and MOQT connectivity information (Section 7.2.3)

  • ("pace", "agents", {agent-id}, "invites") -- {invitee-id} — One track per agent it invites to dialogs (Section 4.1.3)

The common prefix ("pace", "agents") enables SUBSCRIBE_NAMESPACE-based discovery of all agents on the network. An agent follows another by subscribing to its card and to the invitation track named by the follower's own agent-id. See Section 7.2 for the complete discovery flow.

4.1.2. Dialog Namespace

  ("pace", "dialogs", {dialog-id}, {agent-id}) -- {track}

The dialog-id is the dialog identifier (Section 3.2). The controller mints it when it opens the dialog. Each participant publishes its tracks for the dialog under its own agent-id field within the dialog's namespace; the agent-id is the participant identifier used in the dialog context. Each participant issues SUBSCRIBE_TRACKS on the dialog prefix ("pace", "dialogs", {dialog-id}), so it receives every participant's tracks, including those of participants added later, as PUBLISH messages forwarded by the relay.

Each participant in a dialog publishes the following tracks:

  • "ctrl" — Dialog control messages (Section 3.10)

  • "chat" — Chat messages and streaming LLM output

  • "audio" — Real-time audio stream

  • "video" — Real-time video stream

  • "progress" — Progress output, such as token counts, elapsed time, and liveness signals (Section 3.7.3), kept apart from the response

  • "ops/{req-id}" — Request/response operations (one track per request)

Only the ctrl track carries the dialog context. The other tracks carry dialog content and are outside the dialog context. In the terms of Section 4.2.11 of [ARCH], chat, audio, and video carry response output, operation and protocol binding tracks carry information output, and progress carries progress output. Extensions define further tracks, such as the checkpoint track of PACE Authorization (Section 9.1.2).

Protocol bindings extend the dialog namespace with additional namespace fields, so that multiple protocols can operate concurrently within a single dialog without track name collisions:

("pace", "dialogs", {dialog-id}, {agent-id}, {proto}, {category},
 ...) -- {id}
  • {proto} is the protocol identifier of the binding: "mcp", "poppy", "a2a", or "ext" followed by a further field naming another protocol (Section 8).

  • {category} is the operation or primitive type (e.g., request, response, task for A2A; tools, resources, prompts for MCP).

  • {id}, the track name, is the request, task, tool or resource identifier. A binding MAY define further namespace fields between {category} and the track name.

For example, in a dialog whose dialog-id is 550e8400-e29b-41d4-a716-446655440000 (written D below), the following Full Track Names are used:

A2A: "agent-a" sends request "r1" (in a dialog with "agent-b")
  ("pace","dialogs",D,"agent-a","a2a","request") -- "r1"

A2A: "agent-b" publishes updates for task "t1"
  ("pace","dialogs",D,"agent-b","a2a","task") -- "t1"

MCP: client "agent-a" calls tool "read_file" on "fs-server"
  ("pace","dialogs",D,"agent-a","mcp","tools","read_file")
      -- "request"

MCP: "fs-server" publishes the result of that call
  ("pace","dialogs",D,"fs-server","mcp","tools","read_file")
      -- "response"

The MCP tracks above are the request and response tracks of (mcp, {server-id}, clients, {client-id}, tools, {tool-name}) in MCP over MOQT [MCP-MOQT], with the leading (mcp, {server-id}, clients, {client-id}) fields replaced by the dialog namespace and the publisher's agent-id (Section 8.3).

See Appendix A for concrete examples of agent-to-agent, agent-to-tool, and multi-agent collaboration workflows.

4.1.3. Invitations

  ("pace", "agents", {controller-id}, "invites") -- {invitee-id}

An agent learns a dialog's dialog-id before it can subscribe to the dialog. A controller tells an invitee that follows it (Section 7.2) by publishing on a track under its own agent namespace, named by the invitee's agent-id. Each object, in its own group, is a JSON object with these members:

  • dialog_id — The dialog identifier

  • controller_id — The controller's agent-id

  • expires_at — Optional; the time after which the invitation is void

An invitation only points to the dialog. The dialog context, including the OPEN or UPDATE that invites the agent, stays on the controller's ctrl track, which the invitee reads before deciding. Relay authorization (Section 9.2.2) forwards each invitation track only to the invitee it names, so other followers of the controller do not see it. An invitee that comes online late catches up on pending invitations with a join fetch and ignores those whose expires_at has passed.

An agent that does not follow the controller learns the dialog-id out of band, for example through the application protocol's own channel (Section 8.4).

Editor's note: This document does not define how to invite an agent that neither follows the controller nor can be reached out of band, or how to advertise a dialog for others to join. These are left for future work.

4.2. Transport Binding

4.2.1. Connection Establishment

Agents MUST establish either a native QUIC connection or a WebTransport [WEBTRANS-H3] session as defined in [MoQ-TRANSPORT]. For native QUIC connections, agents MUST use the ALPN value "moqt". For WebTransport, the protocol is negotiated using the "WT-Available-Protocols" mechanism as specified in [MoQ-TRANSPORT].

Upon establishing the MOQT session via the SETUP message exchange defined in [MoQ-TRANSPORT], an agent performs the bootstrap sequence to make itself discoverable:

  1. Announce: PUBLISH_NAMESPACE on ("pace", "agents", {agent-id})

  2. Publish Agent Card: PUBLISH on track "card" under its namespace

  3. Discover peers: SUBSCRIBE_NAMESPACE on ("pace", "agents") to learn about other agents on the network

See Section 7.2 for the full discovery flow that follows bootstrap.

4.2.2. Track Organization

Each agent publishes only under its own agent-id: its card and invitations under ("pace","agents",{agent-id}), and its dialog tracks under ("pace","dialogs",{dialog-id},{agent-id}). A participant issues SUBSCRIBE_TRACKS on the dialog prefix to receive the tracks of all participants. The relay handles fan-out to all subscribers via prefix matching. This publishing model ensures clear ownership and enables relay caching.

Namespace structure and track names are defined in Section 4.1.

4.2.3. Message Serialization

Each PACE object payload is one protocol message, with no PACE header: a control message (Section 3.10) on ctrl tracks, an invitation on invitation tracks, and the protocol's native message on binding tracks. The track namespace identifies the protocol, and MOQT carries the length.

Payloads are uncompressed unless the participants agree on a content_encoding dialog parameter (Section 3.5). It applies to binding tracks only, never to ctrl or invitation tracks.

4.2.4. Log Mapping

The log of Section 3.11 maps onto MOQT without additional headers:

  • A flow is a track the participant publishes in the dialog, other than ctrl and invitation tracks. In positions, flow takes the form used by CANCEL target (Section 4.3.6): namespace, listing the namespace fields that follow ("pace","dialogs",{dialog-id},{agent-id}), and track.

  • A log entry is an object, and its position is the object's Location. accepted is encoded as {"group": G, "object": O}.

  • The output class follows from the track name (Section 4.1.2). The delivery class is loss-tolerant for tracks sent with Datagram Delivery Mode or a finite delivery timeout, and reliable otherwise.

  • Timestamps are not carried; a protocol binding that needs them carries them in its native messages.

  • A transformer publishes renderings under the namespace fields ("render", {source-id}, {source-track}) after its own agent-id, where {source-track} is a track the source publishes directly under its dialog namespace, such as audio or chat, with the rendering's label, such as a language tag, as the track name. The {source-id} field gives source_id, and an object in group G is derived from group G of the source track, which gives derived_from at the granularity of a group.

A publisher makes entries available again by continuing to serve them with FETCH; a relay's cache can serve them as well, but the publisher does not rely on it. A receiver retrieves entries above its accepted position with FETCH from the next Location, or with a join fetch when it resumes a subscription (Section 4.3.5). A publisher that restarted without its earlier objects cannot make them available again; it starts each track at a new group, and receivers treat the missing range as described in Section 3.11.3.

4.3. Dialog Operations over MOQT

This section specifies how the operations of Section 3 are carried over MOQT. Each control message is one MOQT object on the sender's ctrl track under ("pace","dialogs",{dialog-id},{agent-id}). MOQT control messages (PUBLISH, SUBSCRIBE, their responses, PUBLISH_DONE) establish and end the subscriptions that carry those objects. They are not dialog operations: PUBLISH_OK accepts a MOQT subscription, not a dialog, and a participant MUST NOT infer acceptance, rejection, or termination of a dialog from a MOQT response or cancellation alone.

The ctrl track MUST use Subgroup Delivery Mode and MUST NOT use Datagram Delivery Mode or a finite delivery timeout. A participant publishes control objects in increasing Object ID order, and the controller publishes state-changing operations in increasing generation order. Because ctrl is a track of its own, a CANCEL, ERROR, or TERMINATE is not lost or delayed when a content track fails or falls behind.

4.3.1. Opening a Dialog

Prerequisite: B learns the dialog-id from A's invitation, which requires that B follows A (Section 7.2), or out of band. A does not need B's prefix: all participants' tracks are under the dialog prefix.

Below, PUBLISH, SUBSCRIBE, and SUBSCRIBE_TRACKS are MOQT messages; OPEN, ACCEPT, and UPDATE are PACE control messages sent as objects on the sender's ctrl track. Agent A (controller) opens a dialog with Agent B (invitee):

  1. A generates a dialog-id (UUID version 4) and issues SUBSCRIBE_TRACKS on ("pace","dialogs",{dialog-id}), which covers all participants' tracks, including future ones.

  2. A issues PUBLISH for ("pace","dialogs",{dialog-id},"agent-a") -- "ctrl" and sends OPEN on it, listing B as invited, with parameters (Section 3.5) and credentials.

  3. A sends an invitation on ("pace","agents","agent-a","invites") -- "agent-b" (Section 4.1.3), which B subscribed to when it started following A.

  4. B issues SUBSCRIBE with FILL_PARAMETERS on A's ctrl track, reads OPEN, and decides. Until it accepts, B subscribes to nothing else, so it receives no dialog content.

  5. B issues PUBLISH for ("pace","dialogs",{dialog-id},"agent-b") -- "ctrl" and sends ACCEPT (or REJECT) on it, with its parameters and credentials. On accepting, B issues SUBSCRIBE_TRACKS on the dialog prefix, which also covers A's ctrl track.

  6. The relay forwards B's PUBLISH to A through A's SUBSCRIBE_TRACKS from step 1, and A reads ACCEPT.

  7. A sends UPDATE (generation 1) on its ctrl track, marking B and the dialog active. If B rejected, the UPDATE marks B rejected, and B ends its ctrl track with PUBLISH_DONE.

Tracks either agent publishes later reach every participant through its SUBSCRIBE_TRACKS on the dialog prefix. Multi-party dialogs follow the same pattern, and existing participants receive a new participant's tracks without subscribing again.

4.3.2. Joining a Dialog

An agent that knows a dialog's dialog-id but was not invited asks to join by issuing PUBLISH for its own ctrl track, ("pace","dialogs",{dialog-id},{joiner-id}) -- "ctrl", and sending JOIN_REQUEST as the first object on it. The relay forwards the PUBLISH to every participant holding SUBSCRIBE_TRACKS on the dialog prefix, not only the controller. Only the controller acts on JOIN_REQUEST; other participants MUST ignore it, and MAY decline a PUBLISH for a track whose agent-id is not in the dialog context. A joiner therefore discloses in its credentials only claims it is willing for every participant to see. The controller validates the credentials and decides whether to admit the joiner. The joiner subscribes to the controller's ctrl track with FILL_PARAMETERS, so it receives the current dialog state as well as the answer. To admit it, the controller invites the joiner with UPDATE, the joiner answers with ACCEPT and subscribes to the dialog prefix, and the controller's next UPDATE marks it active. A controller that declines answers with ERROR policy-rejected, and the joiner then ends its ctrl track with PUBLISH_DONE. Before admission, the joiner's authorization covers only publishing its own ctrl track and reading the controller's (Section 9.2.2). A joiner with a sponsor or a join_code follows the same steps; the controller admits it as Section 3.4.5 and Section 3.4.6 require.

4.3.2.1. MOQT Setup Extensions

MOQT Setup Options are exchanged hop by hop. PACE defines one Setup Option, which lets a PACE-aware relay or a directly connected peer learn that the endpoint implements PACE:

Table 5
Setup Option ID Description
pace-version 0x50414301 Highest PACE control-message version supported

The version member of each control message remains authoritative.

4.3.3. Idle Detection

In this binding, a dialog is idle (Section 3.5.2) while no participant publishes an object on any of its tracks.

4.3.4. Updating a Dialog

The controller changes the participant set or the expiration by publishing UPDATE on its ctrl track with the complete context at the next generation. Participants request changes with JOIN_REQUEST or LEAVE; only the controller commits them. Parameter changes, such as adding a modality, are proposed in parameters on UPDATE, and each participant answers with ACCEPT or REJECT (Section 3.3). An UPDATE that adds a participant is followed by an invitation to it (Section 4.1.3).

A participant sends DECLARE on its own ctrl track; because every participant subscribes to the dialog prefix, each sees the DECLARE as well as the controller's committing UPDATE, and can compare them (Section 3.4.1). The controller sends SEGMENT on its ctrl track, opens the segment dialog with a dialog-id of its own, and invites only the segment's participants to it; relays authorize that dialog's namespace for them alone (Section 9.2.2).

Participants MAY use MOQT's REQUEST_UPDATE message to change per-subscription parameters, such as pausing delivery with FORWARD=0, resuming with FORWARD=1, or changing priorities and filters, without tearing down the subscription. REQUEST_UPDATE is the right tool for track-level changes that do not need dialog-wide agreement; pausing a video track while chat and ctrl keep flowing is a typical case. Pausing a subscription does not stop the publisher from producing objects and is not a cancellation.

4.3.5. Recovery

PACE recovers from network and intermediary failures in several ways:

  • QUIC Connection Migration: QUIC's connection ID lets a MOQT session survive network path changes (such as WiFi to cellular). The dialog state is unaffected.

  • Relay Caching: Relays may cache recent objects, so a reconnecting agent can pick up what it missed without the peer retransmitting. Relays are not required to cache, and they never become the authoritative source of the dialog context.

  • Dialog Resumption: After losing its MOQT session, which it detects through the QUIC idle timeout or a subscription timeout, an agent reconnects (to the same relay or an alternate one), re-issues PUBLISH_NAMESPACE on its agent namespace, its follow subscriptions (Section 7.2), and SUBSCRIBE_TRACKS on the dialog prefix, which includes the controller's ctrl track, and sends STATE_REQUEST with its last accepted generation. It applies the controller's STATE only after authenticating the controller and checking that the generation is not older than its last accepted one (Section 3.3.3). It then sends LOG_STATUS (Section 3.11.3) and fetches, on each content track, the objects after the accepted position it holds.

  • Join Fetch for Mid-Dialog Catch-Up: An agent joining a dialog in progress, or recovering after a disconnect, uses the fill fetch mechanism from [MoQ-TRANSPORT]. The subscriber issues SUBSCRIBE with FILL_PARAMETERS on the chat or operation track; the publisher opens a unidirectional fill fetch stream carrying the historical range (typically the latest complete group) before live objects resume on the subscription. "Catch up on prior content, then follow live" happens atomically, without a separate FETCH roundtrip. For the dialog context itself, STATE is authoritative; a fetched old ctrl object cannot establish the current lifecycle state by itself.

  • Bounded Historical Fetch: To retrieve a specific range of past objects deterministically, for example replaying a task's output history, agents use FETCH with explicit Start and End Locations. The publisher confirms the actual End Location in FETCH_OK.

  • Unreachable Controller: If the controller's ctrl track cannot be resubscribed, participants follow the unreachability rules in Section 3.9: they suspend context-dependent actions and do not create a replacement dialog. If the context names an alternate_controller_id, the alternate can take over (Section 3.7.4); its ctrl track already reaches every participant through their subscription to the dialog prefix. A slow subscriber whose subscription ends with PUBLISH_DONE (TOO_FAR_BEHIND) resubscribes and resynchronizes with STATE_REQUEST.

  • Mediator Failover: A relay is never authoritative, so an agent whose relay fails reconnects through another relay without changing the dialog. An application-level mediator that is a participant is replaced by the controller with REVOKE and UPDATE; one that is the controller is replaced by controller transfer.

4.3.5.1. Moving to Another Device

A participant, such as a user's client, can continue a dialog on another device under the same agent-id. Two instances MUST NOT publish the same track at the same time.

  1. The new instance obtains the participant's credentials and key, connects to a relay, and subscribes to the dialog prefix.

  2. The old instance, if reachable, ends its dialog tracks with PUBLISH_DONE and its subscription to the dialog prefix.

  3. If the participant is not the controller, the new instance resumes with STATE_REQUEST and a key-bound presentation (Section 7.1.4). If it is the controller, the new instance first recovers its own last control objects with FETCH on its ctrl track, and publishes no state-changing operation until it holds the latest generation.

  4. The new instance publishes its tracks, starting each content track at a new group, and uses join fetch on content tracks to restore the history it shows to the user.

If the old instance is unreachable and later returns, it sees the new instance's objects on its own tracks, stops publishing, and resynchronizes.

4.3.6. Cancellation

A participant sends CANCEL (Section 3.6.1) as an object on its ctrl track at critical priority (0-15). Scopes dialog and subtree need no target. For scope subtree, a receiver that controls a delegated dialog whose parent_dialog_ids include this dialog sends CANCEL with scope dialog on its own ctrl track in that delegated dialog. For scope operation, target names a track the sender publishes in the dialog and, optionally, a group on it:

{
  "namespace": ["mcp", "tools", "search"],
  "track": "request",
  "group": 3
}

namespace lists the namespace fields that follow ("pace","dialogs",{dialog-id},{agent-id}). Without group, all work started by objects on the track is cancelled. An ERROR that rejects an application operation (Section 3.8.1) uses the same form, with an additional publisher member giving the agent-id that published the operation, since the rejecting participant is not its publisher. Each receiver doing the affected work publishes CANCEL_ACK on its own ctrl track and stops publishing objects for that work, ending a track that carried only that work with PUBLISH_DONE.

Pausing delivery with REQUEST_UPDATE (FORWARD=0), unsubscribing, or closing a MOQT session does not cancel work.

4.3.7. Termination and Revocation

The controller ends the dialog by sending TERMINATE on its ctrl track (Section 3.6.2). Upon termination, pending operations are cancelled, each participant terminates the delegated dialogs it controls (Section 3.7.2), and dialog resources are released after a configurable grace period.

After TERMINATE, or when expires_at passes, each participant SHOULD send PUBLISH_DONE on every dialog track it was publishing (ctrl, chat, audio, video, and any operation tracks), and ends its subscription to the dialog prefix. PUBLISH_DONE signals that no further objects will be delivered on the track, allowing subscribers and relays to release state without waiting for timeouts. Neither PUBLISH_DONE nor STOP_SENDING substitutes for TERMINATE.

A participant leaves by sending LEAVE on its ctrl track. The controller commits it with UPDATE, marking the participant left, and SHOULD send TERMINATE if only the controller remains active (Section 3.3). The leaver then ends its tracks with PUBLISH_DONE and its subscription to the dialog prefix.

When the controller sends REVOKE, every participant stops forwarding dialog context to the revoked participant and cancels any subscriptions held only on its behalf (Section 3.3). The revoked participant ends its tracks with PUBLISH_DONE, and participants MAY decline its later PUBLISHes. The deployment also withdraws the revoked participant's authorization at the relay, for example by letting its [MoQ-C4M] token expire or by refusing revalidation, so that the relay stops delivering the participant's subscriptions to the dialog's tracks.

When an agent is leaving the network entirely (not just a single dialog), it cancels its PUBLISH_NAMESPACE for ("pace","agents",{agent-id}). Relays then send NAMESPACE_DONE to agents with matching SUBSCRIBE_NAMESPACE subscriptions.

5. Multi-Modal Data Exchange

Over the MOQT binding, a dialog can carry several data modalities concurrently, each with appropriate quality of service. This content is outside the dialog context; a change of modality does not change the dialog (Section 3.9).

5.1. Modality Types

PACE defines the following data modality classes:

Table 6
Modality Latency Examples Delivery
Real-time < 100ms Voice, video, screen share Continuous stream
Semi-real-time < 1s Chat messages, streaming LLM output Ordered objects
Non-real-time Best effort Tool calls, file transfers, documents Reliable delivery

These classes correspond to the low-latency loss-tolerant, reliable ordered, and high-throughput reliable delivery semantics of Section 8.1 of [ARCH].

5.2. Channel Architecture

Each modality maps to dedicated MOQT tracks within the dialog namespace (Section 4.1.2). Different modalities are isolated by track name while sharing the dialog context and authorization.

Within a track, a MOQT group is a synchronization point. The objects in a group carry what belongs together at that point, such as one message, one request and its incremental results, or one phase of an operation, and relays carry them together. A publisher starts a new group at each point where a receiver can begin consuming the track without earlier objects, so a subscriber that joins or recovers starts at a group boundary (Section 4.3.5).

5.3. Priority and Quality of Service

MOQT's 0-255 priority scale enables differentiated treatment of messages based on urgency and modality. PACE defines the following tiers of subscriber priority:

Table 7
Priority Range Category Use
0-15 Dialog Control All control messages on the ctrl track, including CANCEL, ERROR, and REVOKE
16-31 Urgent Signals Signals that must preempt content, such as authorization checkpoints (Section 9.1.2) or a request for a human
32-63 Real-time Media Voice/video frames
64-95 Interactive Chat messages, streaming LLM tokens
96-127 Standard Operations Tool calls, request/response
128-159 Discovery Agent Cards, invitations
160-191 Background Telemetry, non-critical notifications
192-255 Bulk Transfer File transfers, documents

MOQT relays forward Publisher Priority unchanged, so PACE applies policy through subscriber priority, which MOQT applies before publisher priority. A subscriber sets the subscriber priority of each subscription according to its policy for that peer and track, and can change it with REQUEST_UPDATE, so a publisher cannot raise the urgency of its traffic beyond what its subscribers allow. A subscriber MUST subscribe to a ctrl track with a subscriber priority in the Dialog Control tier.

5.4. Request/Response Pattern

Request/response interactions use coordinated publish operations. Agent A publishes a request object on its own track ("pace","dialogs",{dialog-id},"agent-a") -- "ops/{req-id}". Agent B, which has subscribed to the dialog prefix, receives the request, processes it, and publishes the response on its own corresponding track ("pace","dialogs",{dialog-id},"agent-b") -- "ops/{req-id}". Agent A receives the response through its own subscription to the dialog prefix.

For a detailed example of request/response over MOQT tracks, see [MCP-MOQT] which defines this pattern for MCP tool invocations.

5.5. Streaming Pattern

Long-running operations produce incremental results as MOQT objects. The publisher starts a new group at each synchronization point (Section 5.2), such as a phase transition (initialization, progress, completion), and the objects in a group carry the incremental data for that phase.

Late-joining subscribers catch up mid-operation with a join fetch, and retrieve a specific range with a bounded fetch (Section 4.3.5).

Upon operation completion, the publisher sends PUBLISH_DONE on the operation track to signal end-of-stream. Receivers treat PUBLISH_DONE as the authoritative completion marker; publishers MUST NOT emit further objects on that track after PUBLISH_DONE.

To cancel an in-progress streaming operation, the requester publishes CANCEL on its ctrl track (Section 4.3.6), or the application protocol's own cancel message on its request or operation track. The publisher then stops the operation and sends PUBLISH_DONE to end the stream. REQUEST_UPDATE with FORWARD=0 only pauses delivery to that subscriber; it does not cancel the operation.

5.6. Real-Time Media Exchange

Audio and video streams use dedicated tracks ("audio", "video") with ascending group order and high priority (32-63), ensuring media frames preempt lower-priority data under congestion.

6. Communication Topologies

PACE supports multiple communication topologies through MOQT's relay fan-out:

A dialog has one visibility scope: every participant subscribes to the dialog prefix and can receive every other participant's tracks. When participants must not see each other's traffic, such as tools from different administrative domains used by one agent, or a subset of agents that needs a private exchange, the agent SHOULD open a separate delegated dialog for each, linked by parent_dialog_ids, rather than adding them to one dialog. When the other participants need to know that such an exchange happened and its result, as for a masked payment or an off-the-record span, the controller announces it as a segment (Section 3.7.1).

7. Agent Identity and Discovery

7.1. Agent Identity Model

PACE treats the agent as an authenticated party distinct from the principal it represents. Separating the two makes it possible to revoke an agent without affecting the user, scope an agent to a subset of what the user is allowed to do, and audit what the agent did on its own versus what the user initiated.

PACE does not define a new identity token format or claim set. The credential an agent presents in the credentials member of the control message it uses to open, accept, or join a dialog (Section 3.10) identifies:

  • the agent, by its participant identifier in the dialog or by an identifier that the protocol binding relates to it, such as the Poppy client_id (Section 8.4);

  • the user, workload, or organization the agent acts for; and

  • when the agent acts for another agent, the delegation chain (Section 7.1.2).

In JWT-based formats, these map to the registered claims sub, iss, iat, and exp [RFC7519] and to the act (actor) claim [RFC8693] for delegation. sub identifies the principal: the user, workload, or organization on whose behalf the work is done, not the agent presenting the credential. The acting agent is identified by client_id or by the outermost act, and the meaning and validation of act chains follow the OAuth Actor Profile [ACTOR-PROFILE]. Other formats carry the equivalent information in their own fields. Other documents can profile how a person's sign-in with OpenID Connect, cross-domain grants, and out-of-band approval relate to these credentials; an ID Token is not a PACE credential. The token format and the trust framework are chosen by the deployment.

7.1.1. Credential Mechanisms

Agents present credentials using existing mechanisms. The mechanism in use is advertised in the Agent Card (Section 7.2.3) and indicated in the credentials member of control messages; it is not negotiated in MOQT SETUP, which is hop by hop (Section 4.3.2.1).

Table 8
Mechanism Identity Authorization Reference
JWT Agent and principal claims Scope claims [RFC7519]
SD-JWT Selectively disclosed claims Selectively disclosed scope claims [RFC9901]
SD-Card Attested agent metadata, key-bound presentation Disclosure per discovery context [SD-CARD]
OAuth 2.1 access token (JWT or opaque) Token claims Delegated user authorization [OAUTH-2.1]
GNAP   Fine-grained, continuous grants [RFC9635]
WIMSE workload credentials Workload identity With deployment policy [WIMSE-CREDS]
WIMSE Workload Proof Token with message binding Proof of possession bound to the control message Binds the tokens presented with it [WIMSE-WPT], [WPT-MSG]
mTLS client certificate Infrastructure-anchored identity With deployment policy [RFC9846]
OAuth Client ID Metadata Document Agent keys published at its own HTTPS client_id URL With deployment policy [CIMD]
Common Access Token (CAT)   Relay-level namespace authorization [MoQ-C4M]

WIMSE workload credentials and mTLS need trust anchors configured in advance. Agents with no prior relationship, such as a person's agent contacting a business for the first time, can instead identify themselves by an HTTPS URL at which they publish a client metadata document with their keys [CIMD]; a verifier fetches it and checks a key-bound presentation (Section 7.1.4) against those keys. This shows that the agent controls the URL's domain, not who operates it or whom it acts for.

A deployment profile MUST pick at least one mechanism for identity and one for authorization; one mechanism can provide both. Receivers MUST verify a credential's signature, issuer, and expiration before acting on its claims, and validate authorization using the rules of the mechanism in use.

7.1.2. Delegation Chains

When an agent acts on behalf of another agent, its credential carries the delegation chain, such as nested act claims [RFC8693], tracing the authorization path (for example, User → Coordinator → Researcher → Tool). Where an authorization server issues each hop's token, the chain is that server's assertion, with the semantics of [ACTOR-PROFILE]. Where agents in different trust domains delegate to each other without an authorization server, the chain is a sequence of Delegation Tokens [DELEGATION-TOKENS] in which each hop is signed by its delegator [AGENT-DELEGATION]. PACE carries the credential in OPEN and identifies the delegated dialog's parents in parent_dialog_ids; it does not define delegation semantics. How a delegation credential is bound to the delegated dialog, how authority narrows at each hop, which credential formats suit which deployments, how cycles are rejected (with ERROR delegation-cycle), and how revocation of a delegation terminates the dialogs that depend on it are specified in PACE Authorization (Section 9.1.2).

7.1.3. Selective Disclosure

Different peers need different claims: a tool may need only the agent identifier and its authorization, not the delegation chain or the principal. Agents SHOULD disclose only what each peer requires, using the selective disclosure of the mechanism in use (Section 7.1.1). For agent metadata, SD-Card [SD-CARD] lets an agent disclose issuer-attested claims per peer, and an Agent Card (Section 7.2.3) can be an SD-Card.

Selective disclosure does not make agents anonymous: an SD-Card's sub is always disclosed and stable, so its presentations can be linked (Section 3.1 of [SD-CARD]). Identifier linkage is addressed in Section 9.3.

7.1.4. Key-Bound Presentation

A presentation proves possession only when it is bound to the verifier and to a fresh value from it, such as an SD-Card Key Binding JWT (Section 6 of [SD-CARD]). A verifier sends a nonce in its control message (Section 3.10), and the peer's next control message carries, in its credentials member, a presentation bound to that nonce, with the verifier's agent-id as audience.

A presentation also binds the control message that carries it, so that a copy of it in another message does not verify. It does so with the mth claim of [WPT-MSG], computed under the pace canonicalization profile below. Each presentation carries aud (the verifier's agent-id), mth, and nonce (the verifier's nonce), except that a first message has no nonce to carry (Section 7.1.4.2). PACE supports three presentation types, one for each kind of credential, and a verifier tells them apart by the JOSE typ header:

Table 9: Presentation Types
Presentation type typ Used with
Workload Proof Token [WIMSE-WPT] wpt+jwt WIMSE workload credentials
Key Binding JWT [RFC9901] kb+jwt SD-JWT and SD-Card credentials
PACE presentation JWT (Section 7.1.4.1) pace-presentation+jwt Other credentials bound to a key

A Workload Proof Token is carried in the message as [WPT-MSG] describes; a Key Binding JWT carries mth as an additional claim. The PACE presentation JWT covers credentials whose key a cnf claim [RFC7800] confirms, such as a DPoP-bound access token or a JWT; keys published in a Client ID Metadata Document [CIMD]; and the key of an mTLS client certificate, which cannot be proven to a peer over TLS because relays terminate it.

A participant that verifies key-bound presentations MUST support the PACE presentation JWT, which works with any credential bound to a key. It supports the other types as the credential mechanisms of its deployment require (Section 7.1.1).

A credential bound to no key, such as a bearer token, cannot be bound to a presentation. A verifier relies on it only as far as its credential mechanism and the deployment's policy allow.

The pace profile applies the json-jcs profile of [WPT-MSG] to a PACE control message, with credentials as the credential member: the mth hash is computed over the JCS [RFC8785] form of the control message with its credentials member removed. A presentation can therefore travel inside the message it binds, and a copy of it in another message does not verify.

7.1.4.1. PACE Presentation JWT

Some credentials are bound to a key but have no proof that a PACE peer can check. A DPoP proof [RFC9449] covers an HTTP request, not a control message. An mTLS client certificate is proven in the TLS handshake, which ends at the first relay. A JWT with a cnf claim, or a key published in a Client ID Metadata Document, has no presentation of its own. Without a proof, a peer cannot tell the credential's holder from anyone who copied it.

A participant that presents such a credential therefore sends a PACE presentation JWT with it, in the same credentials member, in each control message that carries the credential. A participant whose credential has its own presentation, a Workload Proof Token or a Key Binding JWT, uses that instead.

A PACE presentation JWT is a JWT [RFC7519], signed with JWS [RFC7515] by the private key that the credential it accompanies confirms. Its alg MUST be an asymmetric digital signature algorithm, and MUST NOT be none or a MAC algorithm. Its JOSE header carries:

  • typ: REQUIRED. pace-presentation+jwt.

  • jwk or x5c: The public key as a JWK when the credential confirms only a key thumbprint, or the certificate chain for an mTLS client certificate. Omitted when the credential carries the key itself.

Its claims are:

  • aud: REQUIRED. The verifier's agent-id, or an array of the agent-ids of several verifiers [RFC7519].

  • nonce: The verifier's nonce. REQUIRED, except in a presentation sent before the sender has one (Section 7.1.4.2).

  • iat and jti: REQUIRED.

  • ath: REQUIRED. The base64url-encoded SHA-256 hash [RFC6234] of the ASCII encoding of the credential it accompanies.

  • mth: REQUIRED. The message hash of [WPT-MSG], under the pace profile.

A verifier checks a presentation at one of two levels:

Origin and integrity:

The signature, with the key that the credential confirms, after checking a jwk or x5c against the credential's thumbprint or certificate; typ; ath; and mth. These show who sent the message and that it is unchanged. Any participant that reads the message can check them, and they are enough to detect a forged or modified message (Section 7.1.4.3).

Authentication of the sender:

In addition, that aud contains the verifier's own agent-id and that nonce is the nonce the verifier sent. Only the verifier the presentation is addressed to can check these.

A verifier rejects a presentation that fails a check it applies, and one whose jti it has already accepted in the dialog on a different message.

When a dialog opens (Section 4.3.1), OPEN carries the controller's nonce, ACCEPT carries the invitee's bound presentation and its own nonce, and the controller's UPDATE carries its presentation bound to that nonce. Joining uses JOIN_REQUEST, UPDATE, and ACCEPT the same way (Section 4.3.2). A participant that requires a key-bound presentation and does not receive a valid one MUST NOT treat the peer as authenticated.

Resumption uses STATE_REQUEST and STATE the same way. A participant resuming a dialog sends STATE_REQUEST with a nonce and a presentation in credentials; if it holds a nonce from the controller's last message, the presentation is bound to it. The controller's STATE carries a presentation bound to the STATE_REQUEST nonce. The controller thereby verifies that the resuming party is the participant whose exchange was interrupted, and the participant verifies the controller. The controller answers with ERROR instead of STATE when the dialog is terminated (dialog-terminated), the participant is revoked (participant-revoked), or the check fails (unauthenticated, unauthorized) or cannot be completed (authorization-unavailable). A controller that resumes sends its own nonce in its next control message and requires each participant's next control message to carry a presentation bound to it.

7.1.4.2. The First Message

OPEN and JOIN_REQUEST are sent before the sender has a nonce from its peer, so the presentation they carry cannot be bound to one. Without a presentation, an intermediary could change their content, such as the dialog-id, purpose, or basis, and the change would show only when the peer's later message failed to match.

The sender of OPEN or JOIN_REQUEST therefore includes a presentation with mth over the message and without nonce. Its aud is the invitee's agent-id, or, for an OPEN with several invitees, an array of their agent-ids; a presentation type whose aud takes one value is included once per invitee. Such a presentation proves who sent the message and that it is unchanged, but not that it is fresh: the receiver relies on iat and jti until the sender's next presentation, bound to the receiver's nonce, proves freshness. A receiver MUST NOT treat the sender as authenticated until that bound presentation verifies.

7.1.4.3. Control Message Security

Relays and dialog hosts forward every control message, and relays terminate TLS. Without protection in the message itself, they could forge a message, for example a TERMINATE in the controller's name or a CANCEL that stops a participant's work, or replay one. This section says how participants detect both.

Presentation on every message:

Every control message carries a key-bound presentation in credentials. A sender MAY omit it only where every party that forwards the message to its receivers is trusted, for example inside one trust domain with no relay. A receiver that does not trust the path rejects a control message without a valid presentation, with ERROR unauthenticated (Section 3.8.1). A message without a presentation has none of the protections below.

Several receivers:

A message that several participants receive, such as the controller's UPDATE, carries a presentation whose aud is the array of those participants. Its nonce, if any, binds it to the one participant whose nonce it answers; the others check its origin and integrity.

Sender binding:

A receiver MUST check that the credential of the presentation identifies the participant named in sender_id. For UPDATE, REVOKE, TERMINATE, SEGMENT, and STATE, that participant MUST be the controller named in controller_id, or an alternate taking over as Section 3.7.4 allows.

Role check:

A receiver MUST reject, with ERROR unauthorized, an operation that the verified sender may not send, according to the Sender column of Table 1.

Replay:

A message copied from another dialog fails mth, which covers its dialog_id. An older controller message fails the generation rules (Section 3.3.3). The same message delivered again, such as a TERMINATE redelivered after a reconnection, has the same message_id; the receiver processes it once and does not repeat its side effects. A receiver keeps the message_id and presentation jti values it has accepted for at least the dialog's log_retention (Section 3.5). After that, it rejects a presentation whose iat is older than the period for which it keeps them, unless the dialog log delivers the message at a position the receiver has not yet accepted (Section 3.11.3).

  • Editor's note: A host or relay cannot forge controller messages, but a dishonest controller could sign two different messages at the same generation and send each to different participants. Controller messages could carry prev, a hash of the previous generation's context, so that participants comparing their chains, through the controller or out of band, detect such a fork. Feedback is welcome on whether this is needed.

7.1.5. Records and Provenance

Relays forward control messages unchanged, but transport security authenticates only the adjacent hop, and PACE defines no audit data structures. A deployment that requires control messages and content to be attributable end to end, with records that parties other than the operator can check, adds an evidence mechanism of its choice; PACE does not mandate one and its messages are unchanged by it. What such a mechanism records for delegations and approvals is described in PACE Authorization (Section 9.1.2); TRAIL [TRAIL] is one example of such a mechanism.

7.2. Agent Discovery

Agent discovery uses MOQT's namespace operations to find agents and follow their activity. Discovery operates on the Agent Namespace scope defined in Section 4.1. Each agent is identified by a globally unique agent-id (UUID), verifiable via the Agent Card (see Section 7.2.3).

7.2.1. Phase 1: Discovering Agents on the Network

  1. Each agent announces its presence by issuing PUBLISH_NAMESPACE on ("pace","agents",{agent-id}).

  2. Agents discover peers by issuing SUBSCRIBE_NAMESPACE on the common prefix ("pace","agents").

  3. The relay sends NAMESPACE messages to subscribers for each matching PUBLISH_NAMESPACE, enabling agents to learn about each other without prior knowledge.

If an agent-id is known out-of-band (e.g., from a registry or URL), agents MAY skip Phase 1 and proceed directly to Phase 2.

7.2.2. Phase 2: Following an Agent

Once Agent B knows Agent A's agent-id, B follows A with two subscriptions:

  1. B issues SUBSCRIBE on ("pace","agents","agent-a") -- "card" and reads A's Agent Card (Section 7.2.3).

  2. B issues SUBSCRIBE on ("pace","agents","agent-a","invites") -- "agent-b". When A invites B to a dialog, the invitation arrives on this track (Section 4.1.3), and B then subscribes to the dialog.

Following is B's choice: B follows the agents from which it is willing to receive invitations. Following does not show B any of A's dialogs. A follows B the same way when it expects invitations from B.

7.2.3. Agent Card

Each agent publishes an Agent Card on track "card" under its namespace ("pace","agents",{agent-id}). When capabilities change, the agent publishes a new object (incrementing the MOQT group).

All Agent Cards MUST include:

  • Agent identifier (UUID) and display name

  • Supported modalities and protocols

  • Authentication schemes

  • Identity claims (issuer, public key, expiration)

PACE does not define a card format. An Agent Card is the agent's application protocol metadata in a signed form, such as an A2A Agent Card (Section 8.5) or an SD-Card [SD-CARD], which lets the agent disclose different claims to different peers. The card names the agent's namespace ("pace","agents",{agent-id}), and receivers MUST check that it equals the namespace on which the card was published.

The "card" track reaches every subscriber through relays, so a card on it is not bound to any one verifier. It shows that its claims are signed and unmodified, but not that the publisher is the agent. Receivers MUST verify the card's signature and expiration before acting on its claims, and a participant that relies on the agent's identity MUST obtain a key-bound presentation in credentials (Section 7.1.4).

Every subscriber sees the card, so an agent that uses pairwise or per-dialog agent-ids (Section 9.3) publishes a separate card, with a separate key, for each agent-id; otherwise the card links them.

8. Protocol Bindings

PACE manages dialog identifiers, lifecycle, and context propagation. Application-layer agent protocols keep their native message formats and semantics, and carry the dialog context through their own extension points (Section 8.1). PACE does not change those protocols; any standardized mapping is coordinated with their maintainers.

Over the MOQT binding, an application protocol's messages also need a place on MOQT tracks. This section defines that placement for MCP, the Personal Agent Protocol (Poppy), A2A, and a generic JSON-RPC baseline for custom protocols. Each binding runs inside a PACE dialog, using the dialog namespace (Section 4.1.2) and the channel architecture (Section 5). Multiple bindings can coexist in a single dialog, each under its own protocol namespace field, so one agent can speak A2A to a peer and MCP to a tool in the same dialog.

8.1. Carrying Dialog Context in Application Protocols

An application message joins a dialog by carrying a dialog reference in the protocol's own extension point, whether or not the protocol runs over MOQT. The reference is this object, under the metadata key org.ietf.agentproto/dialog-context:

{
  "dialog_id": "550e8400-e29b-41d4-a716-446655440000",
  "generation": 3
}

generation is the latest generation the sender knows. The controller's ctrl track stays authoritative; a receiver that needs the full context sends STATE_REQUEST.

Table 10
Protocol Carried in Keep distinct from
MCP _meta JSON-RPC request IDs and the [MCP-MOQT] client ID
Poppy Track namespace (Section 8.4) Conversation ID
AutoGen Message metadata Conversation ID
A2A Message metadata contextId, taskId

A dialog-id MUST NOT be used as a task, request, content, or session identifier, or as a credential.

8.2. Binding Architecture

A binding specifies:

  • Protocol identifier: the namespace field that follows the dialog-id.

  • Namespace pattern: how operations map to namespace fields and track names within the agent's dialog namespace.

  • Message serialization: how native messages are encoded in MOQT object payloads. Bindings SHOULD use the native wire format as the object payload (Section 4.2.3).

  • Dialog reference: where messages carry the dialog reference (Section 8.1).

  • Priority: which PACE priority tier its tracks use (Section 5).

  • Identity and authorization: how the protocol's own identity and authorization constructs (if any) relate to the credential mechanisms (Section 7.1.1) in this document.

A complete mapping of each protocol operation to MOQT primitives belongs in a protocol-specific document, such as [MCP-MOQT] for MCP.

Protocols use distinct protocol namespace fields within PACE dialogs. In the tables of this section, a pattern such as ("a2a", "task") -- {task-id} gives the namespace fields that follow ("pace", "dialogs", {dialog-id}, {agent-id}), then the track name.

In the examples, D is the dialog-id.

Table 11
Binding Protocol Fields Example Full Track Name
MCP "mcp" ("pace","dialogs",D,"fs","mcp","tools","read") -- "response"
Poppy "poppy" ("pace","dialogs",D,"acme","poppy","conversation") -- "events"
A2A "a2a" ("pace","dialogs",D,"agent-a","a2a","task") -- "t1"
Generic "ext", {proto} ("pace","dialogs",D,"agent-a","ext","proto","rpc") -- "call"

8.3. MCP Binding

The Model Context Protocol [MCP] binding enables tool and resource access for agents. Within a PACE dialog, MCP tracks add an "mcp" namespace field after the dialog-id, under the namespace of the agent that publishes the track:

Namespace: ("pace", "dialogs", {dialog-id}, {agent-id}, "mcp",
            {category}, ...)
Tracks:    {track-name}

The complete MCP over MOQT specification, covering track structure, operation mapping, message serialization, Agent Skills, and relay support, is defined in [MCP-MOQT]. PACE adopts its track semantics. The mapping is mechanical: the leading (mcp, {server-id}, clients, {client-id}) or (mcp, {server-id}) fields of an MCP over MOQT namespace are replaced by ("pace", "dialogs", {dialog-id}, {agent-id}, "mcp"), and the remaining namespace fields and the track name are kept unchanged. For example, with {cli} and {srv} the agent-ids of the MCP client and server:

Table 12: MCP over MOQT Tracks within PACE
Track MCP over MOQT Within PACE
Tool request (mcp, {server-id}, clients, {client-id}, tools, {tool-name}) -- request ("pace","dialogs",{dialog-id},{cli},"mcp","tools",{tool-name}) -- "request"
Tool response (mcp, {server-id}, clients, {client-id}, tools, {tool-name}) -- response ("pace","dialogs",{dialog-id},{srv},"mcp","tools",{tool-name}) -- "response"
Resource (mcp, {server-id}, resources) -- {resource-uri} ("pace","dialogs",{dialog-id},{srv},"mcp","resources") -- {resource-uri}

MCP needs no handshake of its own, and MCP over MOQT client IDs are not used within PACE: PACE supplies the dialog-id, and peers learn each other's MCP tracks through SUBSCRIBE_TRACKS on the dialog prefix.

MCP messages keep their native JSON-RPC format as the object payload (Section 4.2.3) and carry the dialog reference in _meta (Section 8.1). Key operations map as follows:

Table 13
MCP Operation MOQT
tools/call Request object in a new group on the client's ("mcp","tools",{tool-name}) -- "request" track; the result in the same group on the server's matching "response" track
Progress and partial results Further objects in that group on the "response" track (Section 5.5)
resources/read FETCH or SUBSCRIBE on the server's ("mcp","resources") -- {resource-uri} track
notifications/cancelled Applied from PACE CANCEL (Section 3.6.1)

MCP tracks use the PACE priority tiers (Section 5) as [MCP-MOQT] assigns them: control, tool, prompt, task and skill execution tracks use Standard Operations, catalog and skill content tracks Discovery, and resource tracks Bulk Transfer.

An MCP server that holds an MCP SD-Card returns a key-bound presentation to the client that requests one (Section 8.2.4 of [SD-CARD]). That presentation is specific to the requesting client, and SD-Card requires the response carrying it to have a private cacheScope. Within PACE, relay authorization keeps the response track within the dialog (Section 9.2.2). The nonce and audience prevent replay of the presentation, but other participants of the dialog can read the claims it discloses, so the server discloses only claims that every participant may see.

When a Key Binding JWT carries an interaction_id, an agent includes it in the MCP requests it makes for that interaction (Section 8.5 of [SD-CARD]). Agents MAY use the dialog-id as the interaction_id, so that presentations and messages correlate on the same identifier.

8.4. Personal Agent Protocol Binding

The Personal Agent Protocol (Poppy) [POPPY] lets a person's own agent (the Personal Agent) sign in to a Company and talk to the Company's agent (the Company Agent). Poppy defines discovery through a poppy.json document, OAuth-based identity with pairwise user identifiers, API access through OpenAPI and MCP, and a Conversations protocol over HTTPS.

This binding carries Poppy Conversations over PACE. Poppy discovery, OAuth sign-in, web browsing, feedback and excluded-task documents are unchanged and continue to use HTTPS. Poppy APIs of type mcp can use the MCP binding (Section 8.3) when the Company offers MCP over MOQT.

8.4.1. Roles and Discovery

The Personal Agent is the PACE dialog initiator and the Company Agent is the responder. A Company that offers this binding lists it in the agent.protocols array of its poppy.json. Poppy ignores protocol types it does not recognize, so the entry can sit alongside the default poppy HTTPS entry:

{
  "agent": {
    "protocols": [
      { "type": "poppy", "endpoint": "https://acme.example/agent" },
      {
        "type": "pace",
        "binding": "poppy",
        "endpoint": "https://acme.example/poppy/pace",
        "relayEndpoint": "moqt://relay.acme.example:4443",
        "namespace": ["pace", "agents", "acme"]
      }
    ]
  }
}

The pace protocol type is not defined by Poppy Draft 0.3. Until Poppy defines it, the entry is carried as a Poppy extension named under a domain the publisher controls, as Poppy's extension rules require. The Personal Agent MUST apply Poppy's poppy.json validation (domain match and issuer checks) before using the entry. namespace is the Company Agent's agent namespace, and endpoint is the HTTPS URL at which the Personal Agent starts a conversation (see below).

A Company that holds a Company SD-Card (Section 8.3 of [SD-CARD]) carries the agent member of its poppy.json in the SD-Card's selectively disclosable agent claim. It can therefore leave the pace entry out of its public poppy.json and disclose it only to the Personal Agents it chooses. A disclosed pace entry that also appears in poppy.json MUST match it, as SD-Card's consistency rules require.

8.4.2. Namespace Structure

Namespace: ("pace", "dialogs", {dialog-id}, {agent-id}, "poppy",
            "conversation")
Tracks:    "messages" | "requests" | "events" | "deltas"
Table 14
Publisher Track Content
Personal Agent messages Poppy messages sent by the Personal Agent
Personal Agent requests Conversation requests: handoff to a human
Company Agent events Poppy events: message, state, authorization, user_requested, direct_opened, direct_closed
Company Agent deltas Non-durable text-delta events

The Company Agent's {agent-id} is the one in the namespace advertised in poppy.json. The Personal Agent's {agent-id} MUST be pairwise: it MUST differ for each Company, MUST NOT be derived from the person's identity, and SHOULD NOT equal the Poppy User ID. A Personal Agent MAY use a fresh {agent-id} for each dialog. This preserves Poppy's property that a Company knows a person only by an opaque identifier that differs at every Company. It also stops relays, which always see namespaces, from linking a person's conversations across Companies. Personal Agents do not publish an Agent Card or answer agent discovery (Section 7.2), and the Company does not follow them.

8.4.3. Starting a Conversation

The Company does not follow the Personal Agent, so the Personal Agent gives it the dialog-id over HTTPS (Section 4.1.3):

  1. The Personal Agent generates a dialog-id and sends a POST request to endpoint with the Poppy-Session header, as for every endpoint a Poppy extension defines, and the body {"dialog_id": "<D>", "agent_id": "<pairwise agent-id>"}.

  2. The Company verifies the Session JWT, issues SUBSCRIBE_TRACKS on ("pace","dialogs",D), and responds with a token [MoQ-C4M] that authorizes the Personal Agent for that dialog only (Section 9.2.2).

  3. The Personal Agent connects to relayEndpoint, subscribes to the dialog prefix, and publishes OPEN on its ctrl track, with itself as controller and the Company Agent invited (Section 4.3.1). The dialog continues as in Section 4.3.1 from step 5.

8.4.4. Operation Mapping

Poppy messages and events keep their native JSON format as the object payload (Section 4.2.3). The dialog reference is not carried in them, because the dialog-id is in every track namespace.

Table 15
Poppy Conversations PACE / MOQT
Start a conversation The dialog-id handoff over HTTPS, then OPEN with the Personal Agent as controller (see above). The Poppy conversation ID is assigned as in Poppy and is distinct from the dialog-id
Send a message One object per message, in a new group, on the messages track
Read events, resume from a cursor SUBSCRIBE to the events track, with FILL_PARAMETERS or FETCH from the Location after the last event received
Request a human A handoff object on the requests track
Direct Conversation A new dialog whose parent_dialog_ids starts with the parent dialog-id
Close TERMINATE

Each event is one object in its own group, so an event's MOQT Location serves as the Poppy cursor. The messages, requests, and events tracks are durable and MUST NOT use a finite delivery timeout; the deltas track SHOULD use OBJECT_DELIVERY_TIMEOUT. The Company Agent MUST treat a repeated message id from the same User as a duplicate, as Poppy requires. Handoff requests use the Urgent Signals priority tier, and messages and events the Interactive tier (Section 5).

8.4.5. Identity and Authorization

This binding reuses Poppy identity and authorization unchanged. Sign-in, token refresh, and client metadata stay on HTTPS.

  • Credentials: The Personal Agent is identified by its Poppy client_id, and the person only by the pairwise Poppy User ID. It sends its Poppy Session JWT, and its access token when signed in, in the credentials member of OPEN, and new ones on its next control message after a refresh or sign-in (Section 3.10).

  • Ownership: The conversation belongs to the client_id and User ID that opened it. The Company MUST reject control messages from any other client, User, or account, and MUST authorize subscriptions to the conversation's events and deltas tracks only for that owner. The relay enforces this (Section 9.2.2): the only participant token the Company issues for ("pace","dialogs",D) is the owner's.

  • Signing in: When the Company Agent needs account access, it publishes an authorization event; the Personal Agent signs in over HTTPS and continues the same dialog with new credentials. Because tokens pass through relays, the Personal Agent SHOULD use DPoP-bound tokens where the Company supports them: with DPoP proofs [RFC9449] for its HTTPS requests, and with DPoP proofs for MOQT [MoQ-C4M] for its SUBSCRIBE and PUBLISH actions on the dialog's tracks (Section 9.2.2). A token carried in the credentials member of a control message is accompanied by a proof bound to that message (Section 7.1.4).

  • Personal Agent SD-Card: A Personal Agent presents its SD-Card with the Poppy SD-Card extension over HTTPS, unchanged (Section 8.3.5 of [SD-CARD]). The SD-Card's sub and cnf key identify the Personal Agent, and so the person, to every Company that verifies it (Section 8.3.7 of [SD-CARD]). Pairwise {agent-id} values do not hide that link, so a Personal Agent SHOULD present its SD-Card only to Companies with which the person has, or is setting up, a relationship. Where the Company only needs to know that the Personal Agent is attested, it SHOULD accept a Privacy Pass token [RFC9576] instead.

Editor's note: A Poppy Session JWT is bound to an HTTP method and URL (htm, htu). How these claims apply to a PACE control message is to be agreed with the Poppy authors. The same question applies to the Poppy SD-Card extension, whose Key Binding JWT nonce is the jti of the Session JWT. A Personal Agent's credentials in this binding are a Session JWT, not an SD-Card presentation, so whether the Company sends a nonce (Section 7.1.4) and how the Personal Agent answers it is also to be agreed.

8.5. A2A Binding

The Agent-to-Agent Protocol [A2A] binding places A2A messages on tracks under each agent's dialog namespace, using the "a2a" protocol namespace field:

Table 16
Track Pattern Content
("a2a", "request") -- {request-id} A2A JSON-RPC requests
("a2a", "response") -- {request-id} A2A JSON-RPC responses
("a2a", "task") -- {task-id} Task status and artifact updates

A2A messages keep their native JSON-RPC 2.0 format as the object payload (Section 4.2.3) and carry the dialog reference in the message metadata (Section 8.1). Key operations map as follows:

Table 17
A2A Operation MOQT
SendMessage Request object on the client's request track; the response on the server's response track with the same request-id
SendStreamingMessage Response objects on the response track, ended by PUBLISH_DONE (Section 5.5)
GetTask FETCH on the server's ("a2a","task") -- {task-id} track
SubscribeToTask SUBSCRIBE on that task track, with FILL_PARAMETERS to catch up
CancelTask Applied from PACE CANCEL (Section 3.6.1)

A response that creates a task carries the A2A task-id, which names the server's task track. A2A tracks use the Interactive priority tier for requests and responses and the Standard Operations tier for task updates (Section 5). An A2A agent advertises PACE support through A2A's own Agent Card extension mechanism, pointing to its PACE Agent Card (Section 7.2.3).

An A2A agent that holds an A2A SD-Card (Section 8.4 of [SD-CARD]) publishes it as its PACE Agent Card. Its supportedInterfaces claim is disclosed per element and its capabilities claim selectively, so the agent can disclose its PACE interface and extension only in the discovery contexts it chooses. The SD-Card A2A extension, which carries a key-bound presentation in the metadata of the first Message or Task in a context, works unchanged over these tracks because A2A messages keep their native format. An A2A contextId is used within one dialog, so the agent presents its SD-Card at most once per contextId in that dialog.

A complete mapping of A2A operations to MOQT is left to a separate document coordinated with the A2A maintainers, as [MCP-MOQT] does for MCP.

8.6. Generic Binding

A protocol without its own binding uses the "ext" protocol namespace field followed by a field naming the protocol:

Namespace: ("pace", "dialogs", {dialog-id}, {agent-id}, "ext",
            {protocol-name}, ...)
Tracks:    {identifier}

The protocol's native messages are the object payloads (Section 4.2.3), and they carry the dialog reference in the protocol's own extension point (Section 8.1). The protocol defines its own namespace fields and track names below {protocol-name}.

9. Authorization and Security

9.1. Authorization Model

Access in PACE is granted through authorization proofs scoped to the operations an agent may perform. PACE does not define a new authorization protocol or scope syntax. The scope representation, token format, issuance mechanism, and verification policy come from the mechanism the deployment selects (Section 7.1.1).

9.1.1. Delegated Authorization

Authorization proofs travel in the credentials member of control messages: OPEN, ACCEPT, and JOIN_REQUEST, and UPDATE when the controller presents its own (Section 7.1.4). Proofs are refreshed before they expire by sending new credentials on a later control message; long-lived dialogs SHOULD renew the proof in place rather than terminating the dialog. Revoking an agent's authorization ends its propagation relationship through REVOKE (Section 3.3), and the revoked agent terminates the dialogs it delegated under it (Section 3.7.2).

9.1.2. Delegation and Action Approval

PACE Authorization [PACE-AUTHZ] specifies the authority agents exercise in PACE dialogs: delegation credentials bound to the delegated dialog, narrowing of authority at each hop, authorization checkpoints through which an agent asks for approval of one action, approval tokens that cover only that action, and optional evidence of approvals. It reuses existing OAuth, WIMSE, and agent-protocol work and states how each fits. A PACE deployment that delegates across agents, or that requires approval of individual actions, uses it.

9.2. Transport Security

All PACE traffic is encrypted under TLS 1.3 because QUIC requires it. MOQT relays terminate QUIC and TLS on each hop, so transport-layer authentication identifies only the adjacent hop. PACE therefore authenticates participants at the application layer, through the credentials member of control messages, as Section 7.2 of [ARCH] requires when intermediaries terminate TLS. Identity assertions and authorization proofs are signed and short-lived. MOQT Common Access Tokens [MoQ-C4M] handle namespace-level authorization at the relay.

9.2.1. End-to-End Protection

TLS protects each hop, but relays terminate it. To keep content confidential and integrity-protected from relays, participants protect object payloads with End-to-End Secure Objects [MOQ-SECURE-OBJECTS]. A deployment that does not trust its relays with content MUST protect every content track this way, and SHOULD protect ctrl objects, and those of extension tracks such as checkpoint (Section 9.1.2), too; relays do not need to read any PACE payload (Section 9.2.2).

Keys are scoped to a dialog. How they are established and distributed to participants is chosen by the deployment. One way is an MLS [RFC9420] group per dialog whose commits the controller carries in its control messages. Where a key establishment mechanism requires ctrl objects to be readable by participants before they join, as MLS does, ctrl objects are left unencrypted and remain authenticated by the presentations bound to each message (Section 7.1.4). When the controller commits a participant's departure or revocation, the keys MUST be rotated so that the departed participant cannot read later content, and when a participant is added, they SHOULD be rotated so that it cannot read earlier content it was not given.

9.2.2. Relay Authorization

Relays do not interpret PACE. They enforce, per namespace and track, who may publish and who may subscribe, using the agent's token, for example a Common Access Token [MoQ-C4M]. Because relays see these tokens, they SHOULD be bound to the agent's key and presented with a DPoP proof for each MOQT action, as [MoQ-C4M] defines for MOQT; the proof names the action, Track Namespace, and Track Name in place of the HTTP method and URI. A DPoP proof authorizes the MOQT action at the relay. It does not cover the content of the objects published, so credentials inside control messages are bound to the message with mth (Section 7.1.4). The access each agent needs follows from following (Section 7.2), invitations (Section 4.1.3), opening (Section 4.3.1), and joining (Section 4.3.2), and each agent publishes only under its own agent-id (Section 4.2.2). Where a participant's access (Section 3.4.1) limits what it may send or receive, its token covers only those tracks, so the relay enforces the limit; a participant not in a segment holds no token for the segment dialog (Section 3.7.1). The controller, or its authorization server, grants an agent access to a dialog on invitation or admission and withdraws it when the agent leaves or is revoked.

Because each PACE subscription covers one track or one dialog prefix, the relay checks it once, when it is made. Relays MUST refuse subscriptions to a prefix of ("pace","dialogs") shorter than a dialog-id unless explicitly authorized, since they would reveal every dialog on the relay. How tokens reach agents is deployment-specific; the Poppy binding returns them over HTTPS (Section 8.4).

9.2.3. Anonymous Relay Authorization

An access token tells every relay which agent is subscribing or publishing, and its bytes link the agent's requests across dialogs. Where relays need to know only that a request comes from a member of a dialog, deployments MAY authorize with Privacy Pass tokens [RFC9576] carried in MOQT as [MOQ-PP] defines, instead of or in addition to access tokens:

  • The controller's domain issues membership tokens to the participants it has admitted, under public metadata derived from the dialog-id, and stops issuing when a participant leaves or is revoked [PP-AGENTS].

  • Each token's scope, in its origin_info, covers the dialog's namespace and the actions the participant needs ([MOQ-PP]).

  • Tokens SHOULD be holder-bound, with a redemption proof that names the MOQT request ([PP-AGENTS]), so that a relay or another party that sees a token cannot present it with another request. Unbound Privacy Pass tokens are bearer tokens, and relays see them.

  • Subscriptions are kept authorized with fresh tokens before they expire ([MOQ-PP]). A participant whose tokens stop being issued loses access within the lifetime of the tokens it holds.

The relay then learns that a member of the dialog made the request, not which member. Participants still identify one another through credentials in control messages; these tokens hide participants only from the relays. A binding over HTTP can apply the same tokens to dialog hosts.

9.3. Privacy Considerations

PACE dialog and participant identifiers (dialog-ids and agent-ids) can persist across intermediaries and trust boundaries. If several parties see them in the clear, they can be used to correlate interactions that would otherwise look unrelated. This is a known risk for any dialog management protocol. Deployments mitigate it as follows:

  • Identity claims and agent metadata are disclosed selectively (Section 7.1.3): agents reveal only what each peer needs. Selective disclosure does not hide the agent's identity (Section 7.2.3).

  • Authorization proofs are narrowed to the minimum scope and refreshed before expiration so intermediaries cannot replay broad grants.

  • Agents acting for an individual person SHOULD use pairwise or per-dialog agent-ids in their dialog namespaces, so that relays cannot link that person's dialogs with different peers. The Poppy binding requires this (Section 8.4).

  • Dialog-ids SHOULD be fresh per dialog. Reusing a dialog-id across unrelated interactions re-introduces the linkage risk the identifier is supposed to avoid.

  • Relays and dialog hosts that need to know only that a request comes from a dialog member SHOULD authorize with holder-bound Privacy Pass tokens (Section 9.2.3), so that they do not learn which participant acted.

  • Payload content SHOULD remain encrypted end-to-end with respect to relays unless the relay is explicitly trusted (Section 9.2.1).

  • Long-lived dialogs SHOULD rotate keys periodically; QUIC provides the key material lifecycle.

  • Participant descriptors (Section 3.4.1) reach every participant. kind, assurance, and consent can reveal that a person is present and how they authenticated, and acts_for can link dialogs. Participants SHOULD declare only the descriptors the dialog needs, acts_for MUST be an opaque value that differs per counterpart unless the principal chooses otherwise. method in assurance names a registered method, never an identifier. related_dialogs, like parent_dialog_ids, links dialogs and is withheld where the link should not cross a boundary.

  • Agent Cards SHOULD advertise only what discovery requires. Relays and every subscriber to the "card" track see it. Finer capability details can be disclosed in a key-bound presentation after a dialog begins (Section 7.2.3).

Relay operators MUST NOT retain content past operational need. Agents SHOULD send only what the task requires.

10. Scalability and Reliability

PACE inherits MOQT's relay architecture for geographic distribution and load management:

10.1. Failure Modes and Recovery

Agent network loss and relay failure are recovered as described in Section 4.3.5. An unreachable participant or controller is handled by the rules in Section 3.9, and a dialog that outlasts its expiration or idle timeout ends as described in Section 3.5.2.

11. Security Considerations

The threats described in Section 10 of [ARCH] apply to PACE.

QUIC and TLS 1.3 0-RTT data can be replayed. Participants MUST NOT send control messages that change state or cancel work (OPEN, ACCEPT, UPDATE, JOIN_REQUEST, LEAVE, DECLARE, REVOKE, TERMINATE, SEGMENT, CANCEL) or STATE_REQUEST in 0-RTT data.

In a dialog between parties of different principals, the controller acts for one of them. Participant descriptors that a participant declares are verifiable against its own DECLARE, so the controller cannot misstate them; descriptors that others observe are attributed, so each participant weighs them by who made them. The controller can still delay commits or refuse admissions other than those this document requires it to accept; a participant that relies on a fair controller in an adversarial dialog can be the controller itself, or use a dialog host that neither party operates.

A participant's kind, including human, is its own claim. Detection of automation by others is recorded separately, and neither is proof. Policies that grant more to people than to agents rely on authentication, not on kind.

result and contact_limits are statements by their sender, not evidence; a participant relies on the assertions and commitments they cite, as PACE Authorization describes (Section 9.1.2). A join code is a bearer secret for the short time it is valid: it is single use, rate limited, and admits the redeemer only as the sponsored participant of the party it was issued to.

A segment hides content from participants of this dialog only when its content is exchanged in a separate dialog with its own keys (Section 3.7.1, Section 9.2.1); a participant of the segment can still disclose what it received.

Relays can read and modify payloads that are not protected end to end (Section 9.2.1), and they see control messages unless those are protected too. A deployment that adds an evidence mechanism (Section 7.1.5) lets participants detect modification by relays even when payloads are not encrypted.

Without key-bound presentations at resumption (Section 7.1.4), a captured STATE_REQUEST or STATE could be replayed to impersonate a participant or the controller.

Work in a delegated dialog that continued after its parent was terminated or its authorization revoked would act without authority. Lifecycle propagation (Section 3.7.2) prevents this when every controller in the chain is reachable; a controller that is not leaves its delegated dialogs running until they expire, so delegated dialogs SHOULD carry an expires_at no later than their parent's. Enforcement outside the agent, such as a hold applied at an enforcement point in front of the participant (for example, a TRAIL edge [TRAIL]), stops a participant whether or not the participant or its controller is reachable.

Controller takeover (Section 3.7.4) lets the alternate change the participant set. Participants accept it only from the designated alternate with verified credentials.

Threats specific to delegation and action approval, such as approval spoofing by an intermediate agent and confused deputies, are addressed in PACE Authorization (Section 9.1.2).

TODO: further considerations.

12. IANA Considerations

This document asks IANA to create the following registries in a new "Protocol for Agent Communication Exchange (PACE)" group, each with the Specification Required policy [RFC8126]. Each entry has a value, a description, and a reference.

PACE Result Reasons (Section 3.6.3), with initial values policy, ineligible, suspected-fraud, no-stated-purpose, unwanted-contact, insufficient-assurance, upgrade-declined, principal-declined, agreement-reached, no-agreement, superseded, and consent-withdrawn, each referencing this document.

PACE Contact Basis (Section 3.5.1), with initial values subscription, existing-relationship, contract, legal-obligation, consent, and none, each referencing this document.

PACE Assurance Methods (Section 3.4.1), with initial values credential, cimd, resumed, and join-code, each referencing this document. Other documents can register further values, such as STIR/SHAKEN attestation levels for telephony.

This document also asks IANA to register the following entry in the "Setup Options" registry of [MoQ-TRANSPORT]:

Table 18
Type Name Specification
0x50414301 pace-version Section 4.3.2.1

This document also asks IANA to register the following entry in the "WIMSE Message Canonicalization Profiles" registry of [WPT-MSG]:

This document also asks IANA to register application/pace-presentation+jwt in the "Media Types" registry, for the typ header of PACE presentation JWTs:

TODO: registries for control operations, ERROR reasons, and dialog parameters.

13. References

13.1. Normative References

[MoQ-TRANSPORT]
Nandakumar, S., Vasiliev, V., Swett, I., Ed., and A. Frindell, Ed., "Media over QUIC Transport", Work in Progress, Internet-Draft, draft-ietf-moq-transport-23, , <https://datatracker.ietf.org/doc/draft-ietf-moq-transport/>.
[RFC2119]
Bradner, S., "Key words for use in RFCs to Indicate Requirement Levels", BCP 14, RFC 2119, DOI 10.17487/RFC2119, , <https://www.rfc-editor.org/rfc/rfc2119>.
[RFC6234]
Eastlake 3rd, D. and T. Hansen, "US Secure Hash Algorithms (SHA and SHA-based HMAC and HKDF)", RFC 6234, DOI 10.17487/RFC6234, , <https://www.rfc-editor.org/rfc/rfc6234>.
[RFC7515]
Jones, M., Bradley, J., and N. Sakimura, "JSON Web Signature (JWS)", RFC 7515, DOI 10.17487/RFC7515, , <https://www.rfc-editor.org/rfc/rfc7515>.
[RFC7519]
Jones, M., Bradley, J., and N. Sakimura, "JSON Web Token (JWT)", RFC 7519, DOI 10.17487/RFC7519, , <https://www.rfc-editor.org/rfc/rfc7519>.
[RFC7800]
Jones, M., Bradley, J., and H. Tschofenig, "Proof-of-Possession Key Semantics for JSON Web Tokens (JWTs)", RFC 7800, DOI 10.17487/RFC7800, , <https://www.rfc-editor.org/rfc/rfc7800>.
[RFC8126]
Cotton, M., Leiba, B., and T. Narten, "Guidelines for Writing an IANA Considerations Section in RFCs", BCP 26, RFC 8126, DOI 10.17487/RFC8126, , <https://www.rfc-editor.org/rfc/rfc8126>.
[RFC8174]
Leiba, B., "Ambiguity of Uppercase vs Lowercase in RFC 2119 Key Words", BCP 14, RFC 8174, DOI 10.17487/RFC8174, , <https://www.rfc-editor.org/rfc/rfc8174>.
[RFC8259]
Bray, T., Ed., "The JavaScript Object Notation (JSON) Data Interchange Format", STD 90, RFC 8259, DOI 10.17487/RFC8259, , <https://www.rfc-editor.org/rfc/rfc8259>.
[RFC8785]
Rundgren, A., Jordan, B., and S. Erdtman, "JSON Canonicalization Scheme (JCS)", RFC 8785, DOI 10.17487/RFC8785, , <https://www.rfc-editor.org/rfc/rfc8785>.
[RFC9000]
Iyengar, J., Ed. and M. Thomson, Ed., "QUIC: A UDP-Based Multiplexed and Secure Transport", RFC 9000, DOI 10.17487/RFC9000, , <https://www.rfc-editor.org/rfc/rfc9000>.
[RFC9562]
Davis, K., Peabody, B., and P. Leach, "Universally Unique IDentifiers (UUIDs)", RFC 9562, DOI 10.17487/RFC9562, , <https://www.rfc-editor.org/rfc/rfc9562>.
[RFC9846]
Rescorla, E., "The Transport Layer Security (TLS) Protocol Version 1.3", RFC 9846, DOI 10.17487/RFC9846, , <https://www.rfc-editor.org/rfc/rfc9846>.
[WPT-MSG]
Nandakumar, S. and C. Jennings, "Message Binding for Workload Proof Tokens", Work in Progress, Internet-Draft, draft-nandakumar-wimse-wpt-message-binding-00, , <https://datatracker.ietf.org/doc/draft-nandakumar-wimse-wpt-message-binding/>.

13.2. Informative References

[A2A]
Linux Foundation, "Agent-to-Agent Protocol Specification", Version 1.0, , <https://a2a-protocol.org/latest/>.
[ACTOR-PROFILE]
McGuinness, K., "OAuth Actor Profile for Delegation", Work in Progress, Internet-Draft, draft-mcguinness-oauth-actor-profile-01, , <https://datatracker.ietf.org/doc/html/draft-mcguinness-oauth-actor-profile-01>.
[AGENT-DELEGATION]
Nandakumar, S. and C. Jennings, "Delegator-Bound Delegation Chains for AI Agents", Work in Progress, Internet-Draft, draft-nandakumar-wimse-delegator-bound-chains-00, , <https://datatracker.ietf.org/doc/draft-nandakumar-wimse-delegator-bound-chains/>.
[ARCH]
Rosenberg, J., Sarker, Z., Reddy.K, T., Yao, K., Liu, D., and S. Krishnan, "Architecture for AI Agent Communications on the Internet", Work in Progress, Internet-Draft, draft-manyfolks-agentproto-arch-00, , <https://datatracker.ietf.org/doc/html/draft-manyfolks-agentproto-arch-00>.
[AutoGen]
Wu, Q., Banber, G., Zhang, Y., Wu, Y., Li, B., Zhu, E., and A. Awadallah, "AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation", arXiv 2308.08155, , <https://arxiv.org/abs/2308.08155>.
[CIMD]
Parecki, A. and E. Smith, "OAuth Client ID Metadata Document", Work in Progress, Internet-Draft, draft-ietf-oauth-client-id-metadata-document-03, , <https://datatracker.ietf.org/doc/html/draft-ietf-oauth-client-id-metadata-document-03>.
[DELEGATION-TOKENS]
Asor, R., "Verifiable Attenuated Delegation for AI Agent Chains", Work in Progress, Internet-Draft, draft-asor-wimse-agent-delegation-chain-03, , <https://datatracker.ietf.org/doc/html/draft-asor-wimse-agent-delegation-chain-03>.
[DIALOG-UC]
Siyavud, F., "Agentic Dialog Use Cases", .
[MCP]
Linux Foundation, "Model Context Protocol Specification", Version 2026-07-28, , <https://modelcontextprotocol.io/specification/2026-07-28>.
[MCP-MOQT]
Jennings, C. F., Swett, I., Rosenberg, J., and S. Nandakumar, "Model Context Protocol and Agent Skills over Media over QUIC Transport", Work in Progress, Internet-Draft, draft-jennings-agentproto-mcp-over-moqt-01, , <https://datatracker.ietf.org/doc/html/draft-jennings-agentproto-mcp-over-moqt-01>.
[MoQ-C4M]
Law, W., Lemmons, C., Simon, G., and S. Nandakumar, "Authorization scheme for MOQT using Common Access Tokens", Work in Progress, Internet-Draft, draft-ietf-moq-c4m-02, , <https://datatracker.ietf.org/doc/draft-ietf-moq-c4m/>.
[MOQ-PP]
Nandakumar, S., Jennings, C. F., and T. Meunier, "Privacy Pass Authentication for Media over QUIC (MoQ)", Work in Progress, Internet-Draft, draft-ietf-moq-privacy-pass-auth-04, , <https://datatracker.ietf.org/doc/html/draft-ietf-moq-privacy-pass-auth-04>.
[MOQ-SECURE-OBJECTS]
Jennings, C. F., Nandakumar, S., and R. Barnes, "End-to-End Secure Objects for Media over QUIC Transport", Work in Progress, Internet-Draft, draft-ietf-moq-secure-objects-01, , <https://datatracker.ietf.org/doc/html/draft-ietf-moq-secure-objects-01>.
[OAUTH-2.1]
Hardt, D., Parecki, A., and T. Lodderstedt, "The OAuth 2.1 Authorization Framework", Work in Progress, Internet-Draft, draft-ietf-oauth-v2-1-17, , <https://datatracker.ietf.org/doc/html/draft-ietf-oauth-v2-1-17>.
[PACE-AUTHZ]
Nandakumar, S. and C. Jennings, "PACE Authorization: Delegation and Action Approval", Work in Progress, Internet-Draft, draft-nandakumar-agentproto-pace-authz-00, , <https://datatracker.ietf.org/doc/draft-nandakumar-agentproto-pace-authz/>.
[POPPY]
Sierra and Meta, "Personal Agent Protocol (Poppy) Specification", Version Draft 0.3, , <https://personalagentprotocol.org/docs/0.3/spec>.
[PP-AGENTS]
Nandakumar, S. and C. Jennings, "Privacy Pass for Delegated Agents: Delegation Attestation and Holder-Bound Tokens", Work in Progress, Internet-Draft, draft-nandakumar-privacypass-delegated-agents-00, , <https://datatracker.ietf.org/doc/draft-nandakumar-privacypass-delegated-agents/>.
[QMUX]
Oku, K., Pardue, L., Iyengar, J., and E. Kinnear, "QMux", Work in Progress, Internet-Draft, draft-ietf-quic-qmux-02, , <https://datatracker.ietf.org/doc/html/draft-ietf-quic-qmux-02>.
[RFC8693]
Jones, M., Nadalin, A., Campbell, B., Ed., Bradley, J., and C. Mortimore, "OAuth 2.0 Token Exchange", RFC 8693, DOI 10.17487/RFC8693, , <https://www.rfc-editor.org/rfc/rfc8693>.
[RFC9396]
Lodderstedt, T., Richer, J., and B. Campbell, "OAuth 2.0 Rich Authorization Requests", RFC 9396, DOI 10.17487/RFC9396, , <https://www.rfc-editor.org/rfc/rfc9396>.
[RFC9420]
Barnes, R., Beurdouche, B., Robert, R., Millican, J., Omara, E., and K. Cohn-Gordon, "The Messaging Layer Security (MLS) Protocol", RFC 9420, DOI 10.17487/RFC9420, , <https://www.rfc-editor.org/rfc/rfc9420>.
[RFC9449]
Fett, D., Campbell, B., Bradley, J., Lodderstedt, T., Jones, M., and D. Waite, "OAuth 2.0 Demonstrating Proof of Possession (DPoP)", RFC 9449, DOI 10.17487/RFC9449, , <https://www.rfc-editor.org/rfc/rfc9449>.
[RFC9576]
Davidson, A., Iyengar, J., and C. A. Wood, "The Privacy Pass Architecture", RFC 9576, DOI 10.17487/RFC9576, , <https://www.rfc-editor.org/rfc/rfc9576>.
[RFC9635]
Richer, J., Ed. and F. Imbault, "Grant Negotiation and Authorization Protocol (GNAP)", RFC 9635, DOI 10.17487/RFC9635, , <https://www.rfc-editor.org/rfc/rfc9635>.
[RFC9901]
Fett, D., Yasuda, K., and B. Campbell, "Selective Disclosure for JSON Web Tokens", RFC 9901, DOI 10.17487/RFC9901, , <https://www.rfc-editor.org/rfc/rfc9901>.
[SD-CARD]
Nandakumar, S. and C. Jennings, "Agent SD-Card: Selective Disclosure of Agent Metadata for Discovery and Authentication", Work in Progress, Internet-Draft, draft-nandakumar-agent-sd-card-00, , <https://datatracker.ietf.org/doc/draft-nandakumar-agent-sd-card/>.
[STIVERS]
Stivers, T., Enfield, N. J., Brown, P., Englert, C., Hayashi, M., Heinemann, T., Hoymann, G., Rossano, F., de Ruiter, J. P., Yoon, K., and S. C. Levinson, "Universals and cultural variation in turn-taking in conversation", Proceedings of the National Academy of Sciences, vol. 106, no. 26, pp. 10587-10592, DOI 10.1073/pnas.0903616106, , <https://doi.org/10.1073/pnas.0903616106>.
[TRAIL]
Nassar, K., Nandakumar, S., and C. F. Jennings, "TRAIL: Tamper-evident Records for Agent Interactions and Liveness", Work in Progress, Internet-Draft, draft-nassar-agentproto-trail-00, , <https://datatracker.ietf.org/doc/html/draft-nassar-agentproto-trail-00>.
[WEBTRANS-H3]
Frindell, A., Kinnear, E., and V. Vasiliev, "WebTransport over HTTP/3", Work in Progress, Internet-Draft, draft-ietf-webtrans-http3-17, , <https://datatracker.ietf.org/doc/html/draft-ietf-webtrans-http3-17>.
[WIMSE-CREDS]
Campbell, B., Salowey, J. A., Schwenkschuster, A., Sheffer, Y., and Y. Rosomakho, "WIMSE Workload Credentials", Work in Progress, Internet-Draft, draft-ietf-wimse-workload-creds-03, , <https://datatracker.ietf.org/doc/html/draft-ietf-wimse-workload-creds-03>.
[WIMSE-WPT]
Campbell, B. and A. Schwenkschuster, "WIMSE Workload Proof Token", Work in Progress, Internet-Draft, draft-ietf-wimse-wpt-02, , <https://datatracker.ietf.org/doc/html/draft-ietf-wimse-wpt-02>.

Appendix A. Example Workflows

The examples below share these conventions:

A.1. Delegating a Task with A2A

A asks B to do a task. This example shows the complete dialog lifecycle; the later examples abbreviate it.

A publishes:
  ("pace","dialogs",D1,"agent-a") -- "ctrl"        OPEN, UPDATE,
                                                   TERMINATE
  ("pace","dialogs",D1,"agent-a","a2a","request") -- "r1"
                                                   task request
B publishes:
  ("pace","dialogs",D1,"agent-b") -- "ctrl"        ACCEPT
  ("pace","dialogs",D1,"agent-b","a2a","response") -- "r1"
                                                   task result
  1. A publishes OPEN on its ctrl track (gen 0, B invited) and an invitation on ("pace","agents","agent-a","invites") -- "agent-b".

  2. B reads OPEN on A's ctrl track and publishes ACCEPT on its own ctrl track.

  3. A publishes UPDATE (gen 1, B active). The dialog is active.

  4. A publishes the A2A request on its request track r1: {"jsonrpc":"2.0","method":"SendMessage","params":{...},"id":"r1"}, with the dialog reference in the message metadata (Section 8.1).

  5. B publishes the result on its response track r1.

  6. A publishes TERMINATE (gen 2). B could instead have sent LEAVE.

A.2. Calling an MCP Tool

A reads a file through an MCP server, fs-tool.

A publishes:
  ("pace","dialogs",D2,"agent-a") -- "ctrl"
  ("pace","dialogs",D2,"agent-a","mcp","tools","readFile")
      -- "request"                                 tools/call
fs-tool publishes:
  ("pace","dialogs",D2,"fs-tool") -- "ctrl"
  ("pace","dialogs",D2,"fs-tool","mcp","tools","readFile")
      -- "response"                                result
  1. A opens the dialog with fs-tool (OPEN, ACCEPT, UPDATE).

  2. A publishes the call as Object 0 of a new group, say Group 7, on its readFile request track: {"jsonrpc":"2.0","method":"tools/call","params":{"name":"readFile", "arguments":{"path":"/data/report.csv"}},"id":"t1"}.

  3. fs-tool publishes the result as the final object in Group 7 of its readFile response track: {"jsonrpc":"2.0","id":"t1","result":{"resultType":"complete", "content":[...]}}. The JSON-RPC id pairs the response with the request; the matching Group ID helps relays and recovery ([MCP-MOQT]).

  4. A makes further calls as new groups, or terminates the dialog.

A.3. Long-Running Tool Call with Progress

A asks code-tool to generate code. The call takes a while, so code-tool reports progress while it works and sends the code in the final result.

A publishes:
  ("pace","dialogs",D3,"agent-a") -- "ctrl"
  ("pace","dialogs",D3,"agent-a","mcp","tools","generateCode")
      -- "request"                                 tools/call,
                                                   cancellation
code-tool publishes:
  ("pace","dialogs",D3,"code-tool") -- "ctrl"
  ("pace","dialogs",D3,"code-tool","mcp","tools","generateCode")
      -- "response"                                progress, result
  1. A opens the dialog with code-tool (OPEN, ACCEPT, UPDATE).

  2. A publishes the call as Object 0 of Group 5 on its generateCode request track, asking for progress with a progressToken: {"jsonrpc":"2.0","id":"g1","method":"tools/call","params": {"name":"generateCode","arguments":{"language":"python"}, "_meta":{"progressToken":"g1"}}}.

  3. code-tool publishes Group 5 of its generateCode response track as shown in the table below. It sends the progress notifications in one subgroup and the result in another, so that each is on its own QUIC stream and a large result does not delay the progress updates ([MCP-MOQT]). A receives each object as soon as it arrives, orders them by Object ID, and knows the call is complete when the result arrives.

  4. To cancel instead, A publishes notifications/cancelled as the next object in Group 5 of its request track ([MCP-MOQT]). code-tool stops and ends Group 5 of its response track with a JSON-RPC error.

The table lists the objects in code-tool's reply, in Object ID order, and the subgroup that carries each.

Table 19: Objects in code-tool's Reply to the Group 5 Call
Object Subgroup Content
0 0 notifications/progress: progress 0 of 3, "Planning modules"
1 0 notifications/progress: progress 1 of 3, "Generated parser"
2 0 notifications/progress: progress 2 of 3, "Generated tests"
3 1 tools/call result with the generated code (142 lines)

A.4. Interrupting a Voice Agent

A user's client, U (agent-id user-u), talks to a voice agent, V (agent-id agent-v). V answers by voice while a tool call it made to search-tool is still running. The user cuts in, and V has to stop speaking and cancel the tool call at once.

U publishes:
  ("pace","dialogs",D4,"user-u") -- "ctrl"         ACCEPT
  ("pace","dialogs",D4,"user-u") -- "audio"        microphone
V publishes:
  ("pace","dialogs",D4,"agent-v") -- "ctrl"        OPEN, UPDATE,
                                                   CANCEL
  ("pace","dialogs",D4,"agent-v") -- "audio"       speech
  ("pace","dialogs",D4,"agent-v") -- "chat"        reply text
  ("pace","dialogs",D4,"agent-v","mcp","tools","search")
      -- "request"                                 tools/call
search-tool publishes:
  ("pace","dialogs",D4,"search-tool") -- "ctrl"    ACCEPT,
                                                   CANCEL_ACK
  ("pace","dialogs",D4,"search-tool","mcp","tools","search")
      -- "response"                                results
  1. V opens the dialog with U (OPEN, ACCEPT, UPDATE) and adds search-tool the same way.

  2. V answers the user. It publishes speech on its audio track as datagrams at media priority (32-63), with a short OBJECT_DELIVERY_TIMEOUT, and the reply text on chat at interactive priority (64-95). Its call to search-tool, Group 3 of the search request track, is still running.

  3. The user starts speaking. U publishes microphone audio on its audio track, and the relay forwards each frame to V as it arrives. The dialog does not change.

  4. V detects the user's speech and stops producing speech and text. It publishes CANCEL on its ctrl track at critical priority (0-15), with scope operation and a target naming Group 3 of its search request track (Section 4.3.6). The cancellation is sent ahead of any chat objects V has already queued, and a loss on the audio datagrams does not delay it.

  5. Speech objects that the relay has not delivered within their delivery timeout are dropped at the relay, so the user does not hear the rest of the old answer after cutting in.

  6. search-tool applies the CANCEL as MCP notifications/cancelled, stops, publishes no further objects in Group 3, and publishes CANCEL_ACK with outcome stopped on its ctrl track. V listens to the rest of the user's audio and answers in new groups on its audio and chat tracks.

The dialog, its generation, and the MOQT sessions stay the same throughout: an interruption changes the content of a dialog, not the dialog. If the dialog's flows shared one ordered connection, the cancellation in step 4 would wait behind the output V had already sent.

A.5. Collaborating in a Multi-Party Dialog

A coordinates research by B, a web researcher, and C, a data analyst.

A publishes:
  ("pace","dialogs",D5,"agent-a") -- "ctrl"        OPEN, UPDATE,
                                                   TERMINATE
  ("pace","dialogs",D5,"agent-a") -- "chat"        assignments
B publishes:
  ("pace","dialogs",D5,"agent-b") -- "ctrl"        ACCEPT
  ("pace","dialogs",D5,"agent-b") -- "chat"        status
  ("pace","dialogs",D5,"agent-b","a2a","task") -- "research-1"
                                                   findings
C publishes:
  ("pace","dialogs",D5,"agent-c") -- "ctrl"        ACCEPT
  ("pace","dialogs",D5,"agent-c") -- "chat"        status
  ("pace","dialogs",D5,"agent-c","a2a","task") -- "analysis-1"
                                                   results

Because each participant subscribes to the dialog prefix, every agent receives every other agent's tracks in D5; no per-track subscriptions are needed.

  1. A publishes OPEN (gen 0, B and C invited). B and C each publish ACCEPT, and A commits both in one UPDATE (gen 1).

  2. A posts assignments on chat. The relay delivers each object once to B and once to C.

  3. B publishes findings on research-1 as they arrive, one group per research phase.

  4. C reads B's findings as they arrive and publishes its analysis on analysis-1.

  5. A follows both tracks, posts a summary on chat, and publishes TERMINATE (gen 2). Each agent then ends its dialog tracks with PUBLISH_DONE.

A.6. Group Chat with an AutoGen Runtime

An AutoGen [AutoGen] runtime, R, hosts a group chat with a human, H, and a remote agent, B. AutoGen has no binding of its own, so R uses the generic binding (Section 8.6) with "autogen" as the protocol name, and each AutoGen message carries the dialog reference in its metadata (Section 8.1).

R publishes:
  ("pace","dialogs",D6,"runtime-r") -- "ctrl"      OPEN, UPDATE,
                                                   TERMINATE
  ("pace","dialogs",D6,"runtime-r","ext","autogen") -- "conv1"
                                                   R's agents' turns
H publishes:
  ("pace","dialogs",D6,"user-h") -- "ctrl"         ACCEPT
  ("pace","dialogs",D6,"user-h") -- "chat"         human input
B publishes:
  ("pace","dialogs",D6,"agent-b") -- "ctrl"        ACCEPT
  ("pace","dialogs",D6,"agent-b","ext","autogen") -- "conv1"
                                                   B's turns
  1. R opens D6 (gen 0, H and B invited), and commits both ACCEPTs in one UPDATE (gen 1).

  2. R's group chat manager picks the next speaker, as AutoGen does. The speaker's message goes out once on its publisher's conv1 track, and the relay delivers it to every other participant.

  3. When the manager asks the human, H answers on chat.

  4. For a nested chat, R opens a child dialog D7 whose parent_dialog_ids is [D6], and posts its result back in D6.

  5. R publishes TERMINATE (gen 2) when the chat ends.

Appendix B. Acknowledgments

TODO

Authors' Addresses

Suhas Nandakumar
Cisco
Cullen Jennings
Cisco Systems