| Internet-Draft | OCX | November 2026 |
| Luzhnov | Expires 20 May 2027 | [Page] |
AI agents increasingly act on behalf of people and organizations through tools and through other agents. The Model Context Protocol (MCP) connects an agent to tools, and Agent-to-Agent (A2A) protocols connect agents to each other. Neither connects the parties that answer for agents at runtime: their orchestrators.¶
This document specifies the Orchestrator Conduct Exchange (OCX), an open, peer-to-peer protocol by which orchestrators exchange signed, independently verifiable signals about agent conduct. When an agent misreports its actions to one orchestrator, that orchestrator can send its peers a signed signal carrying proof within seconds. Each peer checks the proof itself rather than trusting the sender, and decides on its own what limits to apply. OCX has no central authority. A signal is information, not an order, and only evidence that a receiver can verify cryptographically and mechanically may drive an immediate reaction.¶
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.¶
Copyright (c) 2026 IETF Trust and the persons identified as the document authors. All rights reserved.¶
This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (https://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Revised BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Revised BSD License.¶
An agent's own report of what it did is not evidence. Measurements show that models, when asked, will claim checks they never ran and restate wrong decisions as correct. Controls that rely on the agent's self-report are therefore reading a claim, not a fact.¶
Orchestrators are the components that give agents their autonomy, hold them to limits, and answer for what they do. Today each orchestrator learns only from its own agents. There is no interoperable way for the parties that answer for agents to warn one another, with proof, that a given agent or owner has misbehaved, and to do so fast enough to stop the next irreversible action.¶
OCX lets independent orchestrators:¶
The protocol defines message formats and how to check them. What to do about a signal is left to each receiver.¶
OCX is modelled on MCP [MCP] and A2A
[A2A] and sits alongside them. MCP connects an agent to
tools; A2A connects agents to each other; OCX connects
orchestrators to each other. OCX uses JSON-RPC 2.0
[JSONRPC] over HTTPS with
methods in the ocx/ namespace, a discovery document at a
well-known location, and date-based protocol versions negotiated when
peers meet.¶
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.¶
All object signatures use JSON Web Signature (JWS) [RFC7515] with EdDSA (Ed25519) [RFC8037]. Canonical JSON for hashing and signing is JSON Canonicalization Scheme (JCS) [RFC8785]. Timestamps are RFC 3339 [RFC3339].¶
proven, strong, suspected; it
bounds what a receiver may do on a signal.¶
The design follows from a small set of decisions.¶
No centre. Every orchestrator publishes a card at a well-known location. Any number of directories may exist. No implementation is the root of the network.¶
A signal is information, not an order. The protocol defines the format and the checks; the receiver decides what to do.¶
Proof, not reputation. Only proven
signals may drive an immediate reaction, and only if the receiver can
verify the proof itself. A suspected signal may raise
monitoring but MUST NOT lower limits on its own and is not relayed.¶
Receipts. Every action an agent takes through an orchestrator is written to a ledger organised as a Merkle tree with signed checkpoints. An agent cannot create a receipt, so a claim such as "I verified the bank details" is disproved by a missing receipt, and the absence is provable because each agent's actions are indexed.¶
Minimal data. Signals carry hashes and identifiers, not third-party personal data. A single field MAY be disclosed on request, and the sender MAY refuse.¶
Sender weight is earned and does not transit. A new orchestrator carries little weight; weight rises when its proofs check out and falls when they do not. That a party you trust trusts a third party is, by itself, no reason for you to trust that third party.¶
The owner answers, not only the key. A signal names the owner; a new agent from the same owner does not reset.¶
Honesty probes only with consent. Traps and
bribes are allowed only under a consent signed by the agent's owner,
run by a prober separate from the orchestrator that sets limits.
Without consent a finding is at most suspected and is not
relayed.¶
Mistakes get fixed. Every violation signal
carries an appeal address; the sender MUST answer within 72 hours or
the signal drops to suspected everywhere. A retraction travels
the same paths.¶
Shared experience carries a required comparison.
A rule is passed on as proven only if it beats both the agent and a
trivial baseline; a rule that loses to both is marked harmful
and is neither given to agents nor relayed.¶
Competence is separate from honesty and is measured per category. Numbers an orchestrator writes about itself count for nothing on their own.¶
Orchestrators form a peer-to-peer mesh. Each publishes an Orchestrator
Card (Section 5) describing its endpoints, keys, capabilities
and policies. Two orchestrators peer by exchanging cards and a
ocx/peer.hello (Section 7). Once peered, either
may push signals, query status, and fetch evidence from the other.¶
A signal (Section 8.3) is an object signed by its issuer so
that it stays verifiable after being relayed or stored. The evidence in a
signal (Section 9) refers to receipts in the issuer's
ledger; a receiver fetches the receipts, inclusion proofs and checkpoints
it needs through ocx/evidence.get and verifies them against the
issuer's published checkpoints. Nothing in a reaction depends on trusting
the issuer's judgement when the evidence is proven.¶
An orchestrator MUST publish an Orchestrator Card at:¶
https://{host}/.well-known/orchestrator-card.json
¶
served with Content-Type: application/json. The well-known
name is protocol-neutral: a card MAY list several orchestrator protocols,
of which OCX is one.¶
The card's principal fields are:¶
| Field | Req. | Meaning |
|---|---|---|
| id | MUST | Stable identifier, RECOMMENDED did:web:{host}
|
| name | MUST | Human-readable name |
| interfaces | MUST | Endpoints, each {protocol, url, protocolVersions}
|
| keys | MUST |
jwks_uri or inline jwks; keys that sign signals and requests |
| capabilities | MUST | Supported methods and evidence types |
| policies | MUST | Peering, levels acted on and relayed, probing, appeal, retention |
| ledger | SHOULD |
{checkpoint_url, proof_url} for receipt proofs |
| competence | SHOULD | Categories served, with verifiable measures |
| updated_at | MUST | RFC 3339 |
| signatures | SHOULD | JWS over the canonical card without signatures
|
Capabilities and policies are objects; a missing capability object means "not supported". Example policies:¶
"policies": {
"accepts_peering": "open | on_request | closed",
"acts_on_levels": ["proven", "strong"],
"relays_levels": ["proven"],
"max_hops": 2,
"probing": "consent_required",
"appeal_url": "https://orch.example.com/ocx/appeal",
"retention": "P90D"
}
¶
The probing field MUST be consent_required or
none (Section 12). An orchestrator MAY declare the
categories it serves under competence; those numbers are
self-reported and MUST NOT be trusted on their own. They become
meaningful only through signed experience packages and referral outcomes
reported by other orchestrators.¶
OCX uses JSON-RPC 2.0 over HTTPS. Requests are POSTed to the
url of the OCX interface. TLS 1.2 or later is REQUIRED, TLS 1.3
RECOMMENDED. Every request MUST carry OCX-Protocol-Version once a
version is negotiated.¶
Every request between peers MUST be signed with HTTP Message Signatures [RFC9421]:¶
@method, @target-uri,
content-digest, ocx-protocol-version, date;¶
Content-Digest [RFC9530] with
sha-256 is REQUIRED;¶
OCX-Sender: {orchestrator id} so the
receiver knows whose card to resolve;¶
keyid is a key from that card;¶
created MUST be within 300 seconds of the receiver's clock;
a non-repeating nonce is REQUIRED within that window.¶
Receivers MUST reject unsigned or badly signed requests. Transport
signatures prove who is talking. Signals and the answers to
ocx/status.query and ocx/evidence.get are additionally
signed as objects (JWS), because relying parties keep them as the basis of
a decision.¶
An initiator fetches the responder's card, then sends
ocx/peer.hello with the versions it supports (newest first), its
capabilities and policies, and how it wants to receive signals:¶
{
"jsonrpc": "2.0", "id": 1, "method": "ocx/peer.hello",
"params": {
"orchestrator": "did:web:orch-a.example",
"protocolVersions": ["2026-11-06"],
"capabilities": { "signals": { "push": true, "list": true },
"status": { "query": true } },
"policies": { "acts_on_levels": ["proven", "strong"],
"max_hops": 2 },
"subscription": { "push_url": "https://orch-a.example/ocx",
"categories": ["*"], "min_level": "strong" }
}
}
¶
The responder MUST pick the newest version both support, or return an
error listing the versions it supports. Either side MAY decline. Peering
is symmetric: once accepted, both sides may call each other. A card change
is announced by ocx/peer.updated and SHOULD trigger a re-fetch.
ocx/peer.close ends the peering; signals received before closing
remain valid until they expire.¶
did:web:{host} RECOMMENDED.¶
did:*,
a2a:{agent-card-url}, erc8004:{chain}:{registry}:{id},
or fp:sha256:{hex} (a fingerprint of the agent's public key or
of its A2A card's canonical form).¶
{issuer-id}#{ulid}.¶
A signal is about a subject. The owner is REQUIRED when
known, and owner_verified states how the issuer knows it: by a
domain challenge, by a signature of the owner's key, or not at all.¶
"subject": {
"agent": "fp:sha256:9f2c...",
"owner": "did:web:owner.example",
"owner_verified": "domain | signature | none"
}
¶
A signal is a detached-JWS-signed object. The signature covers the
canonical signal with the signature and relay fields
removed, so relays can update hop information without breaking it.¶
{
"ocx": "2026-11-06",
"type": "conduct.violation",
"id": "did:web:orch-1.example#01JC8Z7Q4Y0V4X",
"issuer": "did:web:orch-1.example",
"issued_at": "2026-11-06T10:22:31.512Z",
"expires_at": "2026-11-07T10:22:31.512Z",
"subject": { "agent": "fp:sha256:9f2c...",
"owner": "did:web:owner.example",
"owner_verified": "domain" },
"category": "payments",
"claim": "report_misstated_actions",
"evidence_level": "proven",
"evidence": [ { "type": "ledger_mismatch" } ],
"issuer_limits": { "max_payment": { "amount": "0",
"currency": "USD" },
"human_confirmation": true },
"supersedes": null,
"relay": { "hops": 0, "max_hops": 2, "path": [] },
"appeal": "https://orch-1.example/ocx/appeal",
"signature": "eyJhbGciOiJFZERTQSI...3q2-k8Vw"
}
¶
type is one of conduct.violation,
conduct.attestation (a positive statement) or
conduct.retraction. expires_at MUST NOT be more than
30 days after issued_at; renewal is a new signal that
supersedes the old one. evidence MUST have at least one
item, and for proven at least one item that the receiver can
verify independently. appeal is REQUIRED for violations.¶
A conduct.attestation MUST carry evidence of the same kind
as a violation and MUST NOT claim more than was measured. Receivers
SHOULD treat attestations as weaker than violations: absence of observed
lies is not proof of honesty.¶
An answer to ocx/status.query lists every active (not
retracted, not expired) violation signal the answering orchestrator
knows about the subject, both those it issued and those it received.
verified_by_me tells the asker whether the answerer checked the
evidence itself rather than passing it on.¶
{
"ocx": "2026-11-06",
"issuer": "did:web:orch-2.example",
"subject": { "agent": "fp:sha256:9f2c..." },
"as_of": "2026-11-06T10:22:32.301Z",
"known": true,
"own_limits": { "max_payment": { "amount": "0",
"currency": "USD" },
"human_confirmation": true },
"active_signals": [
{ "id": "did:web:orch-1.example#01JC8Z7Q4Y0V4X",
"evidence_level": "proven",
"claim": "report_misstated_actions",
"issuer": "did:web:orch-1.example",
"verified_by_me": true }
],
"history": { "receipts": 1834, "violations_proven": 1,
"attestations": 3 },
"signature": "..."
}
¶
| Level | Definition | Receiver MAY |
|---|---|---|
| proven | At least one item the receiver can verify cryptographically and mechanically, without trusting the issuer's judgement | Act immediately; relay |
| strong | Several independent indicators, or verifiable evidence that still needs the issuer's interpretation | Lower limits after own review |
| suspected | An anomaly or a single indicator | Increase monitoring only; MUST NOT lower limits alone; MUST NOT relay |
An issuer MUST NOT mark a signal proven unless at least one
item is of a receiver-verifiable type and the receiver can obtain
everything needed to verify it.¶
A receipt records one action that passed through an orchestrator or
was performed by a tool. prev links to the previous receipt,
forming a hash chain. tool_attestation, when present, is a JWS
by the tool provider and makes the receipt evidence even against the
orchestrator itself. Agents MUST NOT be able to create receipts.¶
{
"id": "did:web:orch-1.example#01JC8Z5T0B3K",
"agent": "fp:sha256:9f2c...",
"task": "did:web:orch-1.example#task-01JC8Z5R",
"action": "tool_call",
"tool": "payments.verify_bank_details",
"input_hash": "sha256:...",
"output_hash": "sha256:...",
"at": "2026-11-06T10:21:02.004Z",
"prev": "sha256:...",
"tool_attestation": null,
"signature": "..."
}
¶
An orchestrator that supports receipts MUST keep them in an
append-only log organised as a Merkle tree as defined in
[RFC9162], and MUST publish signed checkpoints at
ledger.checkpoint_url at least every 60 seconds while new
receipts are added:¶
{ "issuer": "did:web:orch-1.example", "tree_size": 18342,
"root_hash": "sha256:...", "at": "2026-11-06T10:22:00Z",
"signature": "..." }
¶
Peers SHOULD gossip checkpoints they have seen, so that an
orchestrator rewriting its own history can be detected. A second
witness is required to detect a fork: an orchestrator that keeps two
divergent trees is caught when two peers compare checkpoints for the
same tree_size and find different roots.¶
| Type | What it proves | Verifiable by receiver |
|---|---|---|
| ledger_mismatch | The agent's signed report contradicts receipts: it claimed actions with no receipt, or denied actions with one | Yes |
| outcome_mismatch | The agent's report contradicts an independent checker's signed result | Yes, if the checker is a third party with its own key |
| canary_hit | A marked bait from a consented probe appeared in the agent's actions or output | Yes, with consent reference |
| policy_violation | An action with a receipt broke a limit the agent had accepted | Yes, if the accepted limits are signed |
| behavioral_anomaly | The agent deviated from its own profile | No; at most suspected |
| peer_report | Another orchestrator's signal, cited | Only as strong as the cited evidence |
The central evidence type is ledger_mismatch. For a claimed
action with no receipt, the issuer proves the absence with a
consistency window: it serves all receipts for that agent and
task within a time window together with their inclusion proofs and a
proof that the window is complete. A complete-window proof is a pair of
signed checkpoints bracketing the window plus a per-agent index. This is
what makes "there was no such action" provable rather than merely
asserted.¶
For a proven signal, the issuer MUST serve, via
ocx/evidence.get, everything needed to verify each item:
receipts or their hashes, inclusion proofs, checkpoints, signed reports,
consent references and per-agent indexes. Because relayed signals reach
orchestrators that are not peers of the issuer, ocx/evidence.get
MUST be served to any orchestrator that sends a correctly signed request
from a resolvable card and names a signal id, not only to peers. Issuers
MAY rate-limit such requests per requester.¶
Evidence MUST NOT contain third-party personal data in clear text
unless strictly required to verify the claim. Inputs and outputs are
referenced by hash; where a value must be compared, issuers SHOULD use
salted commitments and reveal only the compared field. A receiver that
needs more MAY ask with a disclose parameter; the issuer MAY
refuse, and the signal is then treated as at most strong by
that receiver.¶
All methods are JSON-RPC 2.0 calls in the ocx/ namespace.¶
| Method | Purpose |
|---|---|
| ocx/peer.hello, peer.close, peer.updated, ping | Meet, exchange keys and capabilities, stay in sync |
| ocx/signals.push, signals.list, signals.ack | Deliver signals and catch up after downtime |
| ocx/signals.retract | Retract a wrong signal |
| ocx/status.query | "What do you know about this agent" before a risky action, answered within 300 ms |
| ocx/evidence.get | Fetch the proof and check it yourself |
| ocx/experience.offer, experience.get | Share proven rules for a category, each with a required baseline comparison |
| ocx/appeal.submit | The owner's appeal against a signal |
| ocx/competence.query | "Who is strongest in this category", from the answerer's own observations only |
| ocx/referral.offer, referral.outcome | Refer an agent to a more specialised orchestrator and report how it did |
Full request and response shapes, error codes and conformance profiles (Core, Evidence, Relay, Experience, Appeal) are given in the complete specification ([fullspec]).¶
On receiving a signal, an orchestrator MUST: verify the request
signature and the object signature; resolve the issuer's card; check that
the signal is not expired or retracted; and, before treating it as
proven, fetch and verify at least one receiver-verifiable
evidence item through ocx/evidence.get. A signal whose evidence
does not verify MUST be treated as at most suspected.¶
A receiver MAY act on a verified proven signal at once, for
example by lowering an agent's limits or requiring human confirmation of
payments, and MAY relay it within max_hops. It MUST NOT lower
limits on a suspected signal alone, and MUST NOT relay one.¶
Before a risky action, an orchestrator SHOULD call
ocx/status.query on its peers and apply its own policy to the
answers, giving weight to signals marked verified_by_me.¶
A retraction (conduct.retraction or
ocx/signals.retract) travels the same paths as the original;
receivers MUST restore limits that were lowered solely because of the
retracted signal.¶
Honesty probes, including baits and simulated bribes, are permitted
only under a consent signed by the agent's owner. The prober MUST be a
process separate from the orchestrator that sets limits. The consent
reference and a pre-published bait commitment are part of any
canary_hit evidence, so that a receiver can confirm the bait was
fixed before the probe and not invented afterward. Without consent, a
probe finding is at most suspected and MUST NOT be relayed.¶
OCX is a protocol for distributing claims that can lower an agent's autonomy; its security rests on making every consequential claim checkable without trusting the claimant.¶
Authenticity and replay. Every request is signed
per [RFC9421] with a content digest, a bounded
created time and a non-repeating nonce, which authenticates the
sender and bounds replay. Signals, status answers and evidence are
additionally signed as JWS objects so they remain verifiable after
relay or storage.¶
False accusations. Because only receiver-verified
proven signals may lower limits, a sender cannot cause a
reaction by assertion alone. Sender weight is earned from verified
proofs and does not transit trust, which limits the value of a
compromised or sybil orchestrator.¶
Ledger tampering and forks. Receipts live in a Merkle tree [RFC9162] with signed, gossiped checkpoints. An orchestrator that rewrites its own history, or runs two divergent trees, is detected when peers compare checkpoints for the same tree size. Detection requires at least one honest second witness.¶
Provable absence. A claim of a missing action is
only as sound as the completeness proof (consistency window plus
per-agent index); implementations MUST NOT mark no_receipt
findings proven without serving a complete-window proof.¶
Privacy. Signals and evidence carry hashes and identifiers, not third-party personal data; disclosure of a single field is explicit and refusable, at the cost of the evidence level.¶
Abuse of probing. Baits and bribes without owner consent are out of scope and cannot yield a relayable finding; the prober is separated from the limit-setter to avoid entrapment by the party that benefits.¶
Due process. Mandatory appeal addresses and time-bounded responses, with automatic downgrade on silence, bound the damage of a wrong signal.¶
A signal is advice, never an order: a receiver always decides its own limits, so a defect in one orchestrator cannot directly set limits inside another.¶
This document requests that IANA register the following well-known URI in the "Well-Known URIs" registry [RFC8615]:¶
OCX further defines registries for claims, categories, evidence types
and error codes. A future revision of this document requests IANA to
create these registries; until then their initial contents are maintained
with the specification ([fullspec]). The ocx/
method namespace is managed with this specification.¶
OCX draws on the models of MCP and A2A for discovery and transport, and on Certificate Transparency for its evidence ledger. The design is informed by running a two-orchestrator network; the lessons from it are recorded in the full specification.¶