Internet-Draft PACE November 2026
Meunier & Wood Expires 19 May 2027 [Page]
Workgroup:
Network Working Group
Internet-Draft:
draft-meunier-webbotauth-capacity-management-00
Published:
Intended Status:
Informational
Expires:
Authors:
T. Meunier
Cloudflare
C. A. Wood
Cloudflare

PACE: Capability Management for Agents

Abstract

This document addresses the noisy neighbour problem without requiring a stable client identifier at the origin. Separately allocated budgets let origins limit consumption beneath a shared platform identity. It describes PACE: a framework for managing budget continuity across requests using a composition of the Web Bot Auth with the Moderation of unLinkable Endorsements (MOLE) architecture [MOLE-ARCH] and protocols [MOLE-PROTOCOLS]. This composition allows for per-client capacity management in a way that allows origins to serve clients above an otherwise constrained public baseline.

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://thibmeu.github.io/draft-capacity-management/draft-meunier-webbotauth-capacity-management.html. Status information for this document may be found at https://datatracker.ietf.org/doc/draft-meunier-webbotauth-capacity-management/.

Source for this draft and an issue tracker can be found at https://github.com/thibmeu/draft-capacity-management.

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 19 May 2027.

▲

Table of Contents

1. Introduction

Agents are increasingly used in business and user workflows, including AI assistants, search indexing, content aggregation, and automated testing. Their relationship with origins is symbiotic. They need information from origins, and origins need signals to apply rate-limit policies. These can be binary, like allow or block when an origin cannot serve traffic, or resource-based to allocate capacity among clients.

Historically, most of these signals have been enshrined into deployments of internet protocols. An agent needs an IP address to receive a response from an origin. Its TLS stack and browser behavior provide further signals. It provides a familiar User-Agent because middleboxes might otherwise filter its traffic (see [USER-AGENT-HISTORY]).

Recent developments break these patterns. Agents use intermediaries to drive traffic, including running workloads in cloud platforms or by using connection proxies ([CONNECT-UDP], [CONNECT-IP], [OHTTP], or plain [HTTP]). As workloads shift to platforms making requests on behalf of users, traditional signals used for moderating traffic are no longer viable.

An address, including an [IPv6] address, is not necessarily a user. Blocking a shared IP range can affect everyone behind it. A platform signing requests under one [WBA-HTTPSIG] identity creates a similar shared fate: when one tenant's agent overloads an origin, limiting the platform also limits its other tenants. [SIGNED-AGENT-METADATA] describes this problem and proposes signed product, project, and session identifiers to let an origin isolate traffic below the platform level.

More granular identifiers are useful when the origin needs to recognize a particular product or project, investigate abuse with its platform, or follow a task within a session. [SIGNED-AGENT-METADATA] advocates for such identifiers. While they are pseudonymous, the receiving origin can link requests within the scope and lifetime of those identifiers. This has several privacy costs:

This presents a natural tension between the needs of an origin and the needs of a user. Identifiers give an origin a way to apply capacity and abuse controls at the granularity provided. This depends on the platform assigning them reliably. Agents using those controls bear the cost of attribution. An origin has to be able to apply policies to traffic whether it finds it abusive or acceptable. Bounding consumption does not always require more fine-grained identifiers.

This document describes PACE, a framework for managing capacity for agents with a combination of Web Bot Auth and MOLE, an architecture for private moderation [MOLE-ARCH], in a way that does not reduce to the shared fate, noisy neighbor problem. In PACE, budgets are allocated to clients through MOLE, where a client can can be a tenant on a platform, a tenant or agent workload, or a user behind any of these, and origins moderate capacity based on those budgets. Traffic is similarly attributed to platforms through Web Bot Auth signatures. Composing these two technologies balances the needs of all parties: origins can establish policies on a per-platform and then per-client basis within that platform, choosing the granularity of client using their anchor set.

Avoiding unnecessary disclosure supports the end-user interests discussed in [INTERNET-FOR-END-USER]. For these capacity decisions, the origin need not receive product, project, or session identifiers. When used, a platform signature remains visible, along with public spending terms and other request data. This approach does not prevent discrimination based on that information.

Such a mechanism is useful in a number of cases. Some examples follow:

  1. An AI agent service that hosts remote agents and concentrates their traffic on a small set of IP addresses. One user's activity can affect others when an origin can only limit the shared address or the platform as a whole.

  2. Network operators carry traffic from device-resident agents behind shared IP ranges. Policies applied to an address range or an entire ASN give these users a shared fate.

  3. VPN providers and enterprise zero trust gateways have a noisy neighbour problem. They may know their users, but should not have to expose them to each origin. An origin able to tell employees apart could single out an executive's activity for instance.

2. Conventions and Definitions

This document uses the HTTP terminology of [HTTP] and the MOLE terminology of [MOLE-ARCH]. The roles can be operated by separate parties or combined.

2.1. Parties and hosting relationships

Platform:

A service that hosts workloads for tenants.

Tenant:

A customer organization, developer account, or billing workspace that runs workloads on a platform, often organized as a project.

Workload:

A deployed application or service, as described in [WIMSE-ARCH]. A software product or service can comprise several workloads.

Agent:

Software that performs tasks, potentially on behalf of a user. A session is a bounded task execution or workflow run, not the agent itself.

User:

A person on whose behalf an agent may act.

A typical hosted deployment has this hierarchy:

Platform
  Tenant
    Workload
      Agent

A tenant runs workloads that execute agents. Users can use several agents, and agents can serve several users or run unattended. Device-resident agents need no hosting tenant.

2.2. MOLE roles and budget scopes

Client:

The MOLE participant that obtains an Endorsement, redeems it for a Credential, and presents that Credential when requesting a resource. An agent or a shared workload component can implement this protocol role.

Anchor:

The party that grants Endorsements under a policy the Moderator accepts. Its relationship with a Client constrains access to new budgets.

Moderator:

The party that issues Credentials and verifies their presentations. It sets the budget policy and may be operated by the origin or on its behalf.

Origin:

The server that controls access to the requested resource.

Intermediary:

A proxy, gateway, or tunnel on the request path. An intermediary that terminates HTTP can sign forwarded requests under its own identity.

A Client may hold several budgets, and an account or workload may receive budgets for several Clients. Those allocations depend on Anchor and Moderator policy. A Credential does not establish a unique account, workload, device, or other principal.

MOLE(X) denotes a presentation for budget scope X. MOLE(C) is a Client's budget, while MOLE(T), MOLE(W), and MOLE(U) denote tenant, workload, and account scopes. These labels do not appear as identifiers in presentations. WBA(Y) denotes a Web Bot Auth signature from intermediary Y.

2.3. Endorsement examples

An accepted Anchor could allocate Endorsements at each scope:

  • Platform: A vetting service grants a hosting provider a finite pool.

  • Tenant: The platform caps a retailer's workspace allocation under its subscription, as in Section 3.

  • Workload: A deployment service authenticates the shopping-assistant workload and caps its allocation across replicas and restarts.

  • Agent: The workload operator allocates a shopping agent's task budget against the workload's allowance. New tasks do not reset that allowance.

  • User: An account provider authenticates a customer and caps allocations across that account's agents. This limits accounts, not unique people.

Clients redeem these Endorsements for Credentials. The Anchor accounts for the allocation scope, which presentations do not reveal. Capacity evidence does not grant permission. Combining budget proofs remains undefined (Section 5.4).

3. Budget limits and abuse

Consider the following scenario. A platform hosts agents for two tenants. Tenant A sends excessive requests while tenant B runs a shopping agent within its allocation. Both tenants' requests carry the same platform signature. Blocking that signature would interrupt both tenants. The origin needs to bound A's consumption and stop abusive requests without requiring either tenant to disclose a stable identifier.

Without per-tenant budgets, the origin sees only the shared platform signature as attribution evidence. If A exhausts the platform's aggregate allowance, the origin also limits B, even though B remains within its intended allocation.

Now suppose the platform also acts as an Anchor accepted by the origin. It grants at most two Endorsements per tenant per epoch. Each Endorsement can be redeemed once at this Moderator for an EC-ACT Credential with ten credits. The Moderator charges one credit per accepted operation attempt and returns an updated Credential with the remaining balance. It returns no credits and offers no top-ups. Credentials expire at the epoch boundary.

In this scenario, credits represent capacity granted by the Moderator under its own access policy. Twenty credits correspond to the capacity allocated to a tenant per epoch. A system based on fingerprints might instead apply a rate limit per IP address, User-Agent, or TLS fingerprint.

Each tenant's allocation therefore permits at most twenty budgeted operation attempts in the epoch. Starting another task does not reset the allowance. This bound depends on the Anchor enforcing the allocation and on other accepted Anchors not supplying additional budgets. The origin verifies each spend without learning which tenant's allocation it consumes.

If the origin detects abuse on A's presented request, its Moderator can record NoUpdate and end that credential chain. The chain is the sequence of Credentials obtained by updating one initial Credential. Its remaining credits are forfeited. B's independently allocated Credentials remain usable. Figure 1 shows this decision. The tenant labels explain the example and are not identifiers carried in presentations.

Platform acting as Anchor At most two Endorsements per tenant Tenant A Tenant B Two 10-credit budgets Two 10-credit budgets Abusive request Allowed request + presentation + presentation Origin and Moderator Both requests carry the platform signature Each presentation spends a separate budget NoUpdate Update (t=0) Presented chain ends Chain continues Remaining credits lost One credit spent
Figure 1: Ending an abusive chain while an independent budget remains usable

This ban applies to the presented chain. It does not revoke A's other Credential or prevent A from obtaining a later allocation. A replacement must count against the tenant's allowance. Exhaustion also prevents further spending from a budget, but an update returning zero credits does not end a chain that still has a sufficient balance. Neither a task boundary nor a tenant boundary is established by the presentation itself.

The platform's aggregate limit and the origin's available capacity still apply to both tenants. Public-baseline requests are outside the illustrated budget. Missing budget evidence does not itself establish abuse.

For the origin to ask the platform to investigate a tenant or change its future allocations, they need a feedback arrangement. One possible extension would require a fresh platform Endorsement before accepting an origin-issued Privacy Pass token, using a reverse-flow exchange. Feedback could also occur out of band. These are possible deployment arrangements, not exchanges specified here. Their binding to the affected allocation, enforcement, and disclosure to each party remain to be defined. They need not require a persistent tenant identifier at the origin.

4. Scope and principles

The goal of PACE is to bound consumption on a per-platform and per-client basis through independently allocated and managed budgets. Exhausting one budget should not exhaust another allocation, provided Anchor issuance and Moderator spending policies keep them separate.

The purpose is to allow origins to offer better public baseline services for its users. Without per-client enforcement, origins are left with overly restrictive baselines (e.g., block everything) because honest clients cannot prove their capacity. In other words, the deployment goal is a baseline level of access without a credential, signature, or some other mechanism that demonstrates capacity. Absence of an active signal does not itself establish abuse [CIVIL].

Permissioned (non-public) resources may require evidence of a right before serving any content. A capability can establish that right without identifying its holder. Where a resource's policy depends on a particular principal and its permissions, authenticating the principal with a stable identifier is appropriate. Capacity evidence alone establishes neither identity nor a right to perform a permissioned operation.

Clients choose which evidence to offer based on their needs and the needs of the origin. Origins should ask only for evidence needed by the requested operation. A verified identity does not establish that a request is harmless, and a valid budget does not grant every permission. Moreover, client evidence should reveal only what the origin needs to apply its policy. Network metadata and other request fields can still identify or correlate clients.

5. Attribution, permission, and capacity

PACE is centered around moderation through attribution and authorization. Specifically, an origin decides whether to serve an operation and how much capacity to allocate to it using evidence presented from the client. This evidence can establish attribution (who is responsible for this traffic?), permission (is the party responsible for this traffic allowed to access the resource?), or capacity (does the party responsible for this traffic have budget for accessing the resource?). A request may need more than one form of evidence.

5.1. Attribution

Attribution addresses the question of who is responsible for traffic, and there are many layers at which this attribution can occur (see Section 2.1). For example, a verified WBA(P) signature attributes its covered components to a key published by P [WBA-HTTPSIG]. The origin can apply an aggregate limit to P without receiving an identifier for each Client behind it. The signature does not establish permission unless origin policy grants it to that signer. A workload can authenticate using a variety of schemes, e.g., with a WIMSE Proof Tokens, HTTP message signatures, or even transport-layer authentication (see [WIMSE-AIMS]). Users can authentiate using [OIDC] or related techniques.

5.2. Permission

Permission addresses the question of whether the attributed party can access the resource in question. As such, permission depends on the resource's policy. For example, an OAuth access token can convey an authorization grant [OAUTH]. The origin can check the token as its authorization server requires, including its audience, validity, permissions, and any proof of possession. Identity obtained through [OIDC] or workload authentication can also be an input to access policy.

A workload can obtain a token using [WAG], and an existing federation can support delegated access through [OAUTH-XAA]. These grants are exchanged at the authorization server, which decides what access token to issue. They do not define a common token claim layout. [AAUTH] defines a separate signed authorization protocol with its own challenges and credentials.

Notably, authentication alone does not determine the permitted operations. In fact, an origin can accept a capability without identifying its holder, including a Privacy Pass token when the issuer's attestation satisfies that policy [PRIVACYPASS].

5.3. Capacity

Capacity addresses the question of whether the attributed party has available budget for accessing a resource. Budget can be allocated and managed using granular identities such as a session ID, user ID, or platform ID. Budget can also be allocated and managed using MOLE, which permits clients to hold a hidden balance c and then update it through interaction with the origin and its moderator. (This is referred to as presenting a MOLE credential.) The origin learns the public amounts and that the spend was possible, without learning the balance or a stable client identifier [MOLE-PROTOCOLS].

For example, a policy can issue B credits, charge one per accepted presentation, and return zero with no top-ups. Each issuance then allows at most B fresh accepted operation attempts. Further issuances increase that allocation. Bounding consumption over time also requires policy for replenishment and expiry. Refunding every spend in full leaves successful consumption unbounded.

5.4. Composition

Attribution, authorization, and capacity evidence can compose in requests. For instance, an origin might require that requests carry a platform signature (WBA(P)) and a MOLE presentation corresponding to the user (MOLE(U)) on each request. An origin might also require both a workload authorization token granting permission to use a resource and a MOLE presentation for the user behind that workload. Origins decide what composition is essential based on their needs.

Carriage of a MOLE presentation alongside another authorization scheme or additional budget presentations, and signaling cumulative requirements, remain undefined here (see issue #9).

6. Moderating consumption

Given mechanisms for establishing attribution, permission, and capacity, origins can moderate composition accordingly. The origin's moderator configures its accepted Anchor Set, credential contexts, allocation and spending policy, and any permission requirements. A client may send an initial request without evidence. The origin then verifies what is present and serves the level of access its policy permits. For example, a public response can offer more capacity through a MOLE challenge. A permissioned resource can refuse the operation and challenge for the missing authorization.

A MOLE redemption challenge lets a client obtain a credential. A presentation challenge specifies its spending terms. Clients use the exchanges and verification rules of [MOLE-PROTOCOLS] and [MOLE-HTTP], including replay checks and request binding. A failed proof supplies no evidence of budget. A refund replaces the spent Credential with an updated one. Returning zero credits leaves c - s credits. Recording NoUpdate ends the chain, as illustrated in Section 3. The Moderator records this decision before attempting the operation, using the evidence available at that point.

After a lost response, the Client can recover the recorded credential result within the retention period. Recovery does not repeat the protected operation or establish whether it completed. Application recovery remains separate. One credential permits one presentation in flight at a time.

6.1. Public resource

The origin offers a budget above its public baseline. Client C already holds an Endorsement from an accepted Anchor. Intermediary P adds its own signature. The Moderator is shown at the origin.

Anchor Client C Intermediary Origin P Moderator Endorsement request + WBA(P) baseline + redemption and presentation challenges Redeem & Issue Credential with balance c request + MOLE(C) + WBA(P) record spend, attempt operation response with Update Client balance: c - s + t
Figure 2: Public access with carrier attribution and an anonymous budget
  1. The Client sends a request. The intermediary adds WBA(P), which the origin verifies before applying its aggregate limit.

  2. The origin serves the public baseline and offers redemption and presentation challenges on the 200 response. A Client with a suitable Credential skips redemption. Otherwise, it fetches the Moderator configuration and redeems its Endorsement at the Redeem & Issue endpoint.

  3. The Client presents proof that its balance covers s and requests a top-up of at most a. The Moderator verifies and records the spend result atomically before attempting the operation.

  4. The response returns an Update with 0 <= t <= s + a, continuing the chain with balance c - s + t. Each subsequent spend answers an applicable challenge and uses a fresh nullifier, without a credential link to the previous presentation.

A direct Client uses the same exchange without WBA(P). A Client unable or unwilling to present continues at the public baseline, subject to origin policy. A failed presentation does not authorize the budgeted level of access.

7. Deploying PACE

A deployment of PACE admits two kinds of capacity control: (1) through an intermediary which identify itself for an aggregate, platform-specific limit, and (2) through smaller, anonymous budgets that are allocated to indivdual clients (tenants, workloads, agents, or users) based on the chosen anchor set. Figure 3 shows one example where an Anchor endorses Client C while intermediary platform P carries its requests. The same provider can operate both the role of anchor and intermediary, but need not.

Anchor Endorsement Client C MOLE(C) Intermediary MOLE(C) Origin P Moderator +WBA(P)
Figure 3: Separate endorsement and forwarding relationships

The origin can attribute the forwarded request to P through WBA(P) and charge C's separate budget through MOLE(C), without learning C's identity from that presentation. The allocation relationship, rather than P's signature, establishes which smaller unit bears the spend. The following subsection applies this separation to product, project, and session capacity limits.

7.1. Capacity scopes

Using the terms in Section 2.1, a platform P hosts tenants T with workloads W. An account U may authorize several agents. Section 2.3 describes allocations at these scopes. The noisy-neighbour boundary depends on which of these shares a budget. A tenant budget separates tenants but leaves workloads within a tenant sharing its allocation. A workload budget separates workloads but still groups their Clients. Account budgets can isolate those allocations further, while several Clients sharing one account budget remain subject to the same limit.

In Figure 4, Anchor W endorses workload-scoped allocations and Anchor U endorses account-scoped allocations. Client components obtain and spend the corresponding Credentials. The request carries both budget proofs alongside the platform's signature. The Anchors need not operate the forwarding intermediary. These are intended scopes, pending the multiple-presentation carriage in Section 5.4.

Anchor W Anchor U Endorsement Endorsement Platform P Tenant workload with Clients Intermediary MOLE(W) MOLE(U) +WBA(P) Origin Moderator
Figure 4: Workload and account budget scopes in a proposed stacked deployment

7.2. Composing scopes

A client can receive several independent budgets, for example one per tenant, workload, agent, or user. Authorization determines what each client may do, while issuance policy constrains their combined allocation. Adding Clients must not yield unbounded replacement budgets. Aggregate platform limits should still apply when individual budgets remain available.

With MOLE, concurrent requests need separate credentials, since each request effectively "locks" a credential while in use. Several Credentials provide separate balances, so assigning more Credentials can increase both concurrency and total consumption. Shared-budget splitting and simultaneous checks of aggregate and individual anonymous budgets remain open.

A CDN can operate a Moderator for several origins. All servers accepting the same credential key and context must share atomic replay state. Independent stores need disjoint keys or contexts. Sharing a credential context also shares its budget scope and can enlarge the anonymity set, but activity at one origin can then reduce the balance available at another.

Anchors constrain new allocations, but the Moderator cannot identify the issuing Anchor from a redemption. [MOLE-ARCH] leaves privacy-preserving Anchor feedback unresolved. Removing an Anchor requires evidence from outside that anonymous redemption, and changing sets to probe Clients can reduce privacy. Endorsements from several accepted Anchors can yield several budgets.

Deployment also needs bounded verification work and replay storage. Invalid proofs can consume verification capacity, and replay markers must remain while their Credentials are accepted. Aggregate controls remain necessary for availability, even though they can reintroduce shared fate.

8. Security considerations

The security goals and limitations of MOLE are described in [MOLE-ARCH], [MOLE-PROTOCOLS], and [EC-ACT].

8.1. Request integrity

An intermediary that sees a presentation can race the Client to spend it. The request-bound profile of [MOLE-PROTOCOLS] must cover every component that affects authorization, including the body where relevant. Single-use verification alone determines only which spend wins. Binding does not prevent an intermediary from withholding or racing an identical request.

A carrier signature should cover the MOLE presentation and the relevant request components. It attributes that pairing to the carrier. Protection against the carrier itself requires binding the Client's evidence to the operation. This document does not define the intermediary signing profile.

Sender-constrained access tokens limit reuse of stolen tokens, but DPoP does not protect the request body [DPOP]. Mutual TLS authenticates the connection at its TLS endpoint [MTLS], which may be an intermediary. Application-layer signatures can preserve integrity across TLS termination [WIMSE-AIMS]. For signed HTTP content, the verifier checks Content-Digest against the received content and verifies a signature covering it, together with the other components needed to bind the operation [HTTPSIG] [DIGEST].

8.2. Correlation and policy

Unlinkable Credentials do not hide addresses, cookies, content, or timing. A forwarding provider that knows the Client can disclose its requests to the origin. This deployment does not prevent that collusion.

Credential keys, contexts, and public amounts partition the anonymity set. They should follow coarse policies shared by many Clients. Per-Client values or repeated balance probes can make presentations distinguishable. The privacy requirements of [MOLE-PROTOCOLS] apply to allocation, spending, and refund policy. A platform signature further narrows the observable set.

A Moderator can end a credential chain without identifying its holder. Clients therefore rely on it to apply the advertised policy. Withholding refunds can consume scarce Endorsements, while recovery distinguishes a lost response from a recorded termination only while the result remains available.

9. IANA considerations

This document has no IANA actions.

10. Informative References

[AAP]
"Agent Attestation Protocol", n.d., <https://github.com/agentic-protocols/agentic-attestation-protocol/blob/1328a6bbfda24ea18bcd8996a5417a6702c8f274/draft-agent-attestation-protocol.md>.
[AAUTH]
Hardt, D., "AAuth Protocol", Work in Progress, Internet-Draft, draft-hardt-oauth-aauth-protocol-11, , <https://datatracker.ietf.org/doc/html/draft-hardt-oauth-aauth-protocol-11>.
[BEARER]
Jones, M. and D. Hardt, "The OAuth 2.0 Authorization Framework: Bearer Token Usage", RFC 6750, DOI 10.17487/RFC6750, , <https://www.rfc-editor.org/rfc/rfc6750>.
[CIVIL]
Meunier, T. and K. Kelley, "Principles for Web Traffic Management by Intermediaries", Work in Progress, Internet-Draft, draft-meunier-bcp-civil-00, , <https://datatracker.ietf.org/doc/html/draft-meunier-bcp-civil-00>.
[CONNECT-IP]
Pauly, T., Ed., Schinazi, D., Chernyakhovsky, A., Kühlewind, M., and M. Westerlund, "Proxying IP in HTTP", RFC 9484, DOI 10.17487/RFC9484, , <https://www.rfc-editor.org/rfc/rfc9484>.
[CONNECT-UDP]
Schinazi, D., "Proxying UDP in HTTP", RFC 9298, DOI 10.17487/RFC9298, , <https://www.rfc-editor.org/rfc/rfc9298>.
[DIGEST]
Polli, R. and L. Pardue, "Digest Fields", RFC 9530, DOI 10.17487/RFC9530, , <https://www.rfc-editor.org/rfc/rfc9530>.
[DPOP]
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>.
[EC-ACT]
Schlesinger, S., Katz, J., and A. F. Faz-Hernandez, "EC-ACT: Anonymous Credit Tokens", Work in Progress, Internet-Draft, draft-schlesinger-mole-ec-act-00, , <https://datatracker.ietf.org/doc/html/draft-schlesinger-mole-ec-act-00>.
[HTTP]
Fielding, R., Ed., Nottingham, M., Ed., and J. Reschke, Ed., "HTTP Semantics", STD 97, RFC 9110, DOI 10.17487/RFC9110, , <https://www.rfc-editor.org/rfc/rfc9110>.
[HTTPSIG]
Backman, A., Ed., Richer, J., Ed., and M. Sporny, "HTTP Message Signatures", RFC 9421, DOI 10.17487/RFC9421, , <https://www.rfc-editor.org/rfc/rfc9421>.
[INTERNET-FOR-END-USER]
Nottingham, M., "The Internet is for End Users", RFC 8890, DOI 10.17487/RFC8890, , <https://www.rfc-editor.org/rfc/rfc8890>.
[IPv6]
Deering, S. and R. Hinden, "Internet Protocol, Version 6 (IPv6) Specification", STD 86, RFC 8200, DOI 10.17487/RFC8200, , <https://www.rfc-editor.org/rfc/rfc8200>.
[MOLE-ARCH]
Schlesinger, S., Jackson, D., and T. Meunier, "Moderation of unLinkable Endorsements (MOLE) Architecture", Work in Progress, Internet-Draft, draft-jms-mole-architecture-01, , <https://datatracker.ietf.org/doc/html/draft-jms-mole-architecture-01>.
[MOLE-HTTP]
Schlesinger, S., Jackson, D., and T. Meunier, "MOLE HTTP Transport", Work in Progress, Internet-Draft, draft-jms-mole-http-transport-01, , <https://datatracker.ietf.org/doc/html/draft-jms-mole-http-transport-01>.
[MOLE-PROTOCOLS]
Schlesinger, S., Jackson, D., and T. Meunier, "MOLE Protocols", Work in Progress, Internet-Draft, draft-jms-mole-protocols-01, , <https://datatracker.ietf.org/doc/html/draft-jms-mole-protocols-01>.
[MTLS]
Campbell, B., Bradley, J., Sakimura, N., and T. Lodderstedt, "OAuth 2.0 Mutual-TLS Client Authentication and Certificate-Bound Access Tokens", RFC 8705, DOI 10.17487/RFC8705, , <https://www.rfc-editor.org/rfc/rfc8705>.
[OAUTH]
Hardt, D., Ed., "The OAuth 2.0 Authorization Framework", RFC 6749, DOI 10.17487/RFC6749, , <https://www.rfc-editor.org/rfc/rfc6749>.
[OAUTH-XAA]
Parecki, A., McGuinness, K., and B. Campbell, "Identity Assertion JWT Authorization Grant", Work in Progress, Internet-Draft, draft-ietf-oauth-identity-assertion-authz-grant-06, , <https://datatracker.ietf.org/doc/html/draft-ietf-oauth-identity-assertion-authz-grant-06>.
[OHTTP]
Thomson, M. and C. A. Wood, "Oblivious HTTP", RFC 9458, DOI 10.17487/RFC9458, , <https://www.rfc-editor.org/rfc/rfc9458>.
[OIDC]
"OpenID Connect Core 1.0", n.d., <https://openid.net/specs/openid-connect-core-1_0.html>.
[POPPY]
"Personal Agent Protocol, draft 0.3", n.d., <https://github.com/personal-agent-protocol/personal-agent-protocol/blob/1de1990e4978b42179fe0e3e0ae554271808c308/content/versions/0.3/spec.mdx>.
[PRIVACY-PARTITIONING]
Kühlewind, M., Pauly, T., and C. A. Wood, "Partitioning as an Architecture for Privacy", RFC 9614, DOI 10.17487/RFC9614, , <https://www.rfc-editor.org/rfc/rfc9614>.
[PRIVACYPASS]
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>.
[PRIVATETOKEN]
Pauly, T., Valdez, S., and C. A. Wood, "The Privacy Pass HTTP Authentication Scheme", RFC 9577, DOI 10.17487/RFC9577, , <https://www.rfc-editor.org/rfc/rfc9577>.
[SIGNED-AGENT-METADATA]
Vilgalys, R. and P. González, "The Signed-Agent-Metadata HTTP Header Field for Web Bot Auth", Work in Progress, Internet-Draft, draft-vilgalys-webbotauth-signed-agent-metadata-00, , <https://datatracker.ietf.org/doc/html/draft-vilgalys-webbotauth-signed-agent-metadata-00>.
[USER-AGENT-HISTORY]
"WebAIM: History of the browser user-agent string", n.d., <https://webaim.org/blog/user-agent-string-history/>.
[WAG]
Carleton, P., Steele, N., Parecki, A., Schwenkschuster, A., and B. Campbell, "Workload Authorization Grant", Work in Progress, Internet-Draft, draft-carleton-workload-authz-grant-02, , <https://datatracker.ietf.org/doc/html/draft-carleton-workload-authz-grant-02>.
[WBA-ABA]
Rescorla, E. and R. Barnes, "Anonymous Bot Authentication: Authorization and Rate Limiting for Web Agents", Work in Progress, Internet-Draft, draft-rescorla-anonymous-webbotauth-01, , <https://datatracker.ietf.org/doc/html/draft-rescorla-anonymous-webbotauth-01>.
[WBA-HTTPSIG]
Meunier, T. and S. Major, "HTTP Message Signatures for automated traffic", Work in Progress, Internet-Draft, draft-ietf-webbotauth-httpsig-protocol-01, , <https://datatracker.ietf.org/doc/html/draft-ietf-webbotauth-httpsig-protocol-01>.
[WIMSE-AIMS]
Kasselman, P., Lombardo, J., Rosomakho, Y., Campbell, B., Steele, N., and A. Parecki, "AI Identity Management System", Work in Progress, Internet-Draft, draft-ietf-wimse-aims-01, , <https://datatracker.ietf.org/doc/html/draft-ietf-wimse-aims-01>.
[WIMSE-ARCH]
Salowey, J. A., Rosomakho, Y., and H. Tschofenig, "Workload Identity in a Multi System Environment (WIMSE) Architecture", Work in Progress, Internet-Draft, draft-ietf-wimse-arch-08, , <https://datatracker.ietf.org/doc/html/draft-ietf-wimse-arch-08>.

Appendix A. Acknowledgments

The authors would like to thank Bryan Becker, Marwan Fayed, and Mari Galicer for discussions that helped shape this document.

Appendix B. Capacity without identifiers

Signed product, project, and session IDs give an origin visible keys with which to keep separate counters [SIGNED-AGENT-METADATA]. For capacity alone, the Anchor can constrain Endorsements per product, project, or task allocation only if it knows and enforces those boundaries. The platform can play this role. An independent Anchor needs a way to account for the same allocations, or its Endorsements cannot enforce the platform's project and product totals.

Clients redeem those endorsements for independent MOLE credentials and spend the credentials on requests. The Moderator checks each spend and carries the remaining balance into an unlinkable replacement Credential. The presentation itself does not reveal which product, project, task, or Client holds a particular budget.

The platform can still sign WBA(P), so the origin can also limit its traffic in aggregate. Credential contexts and spending terms should be shared across many endorsements. Unique contexts or amounts would distinguish those units and shrink their anonymity sets.

The allocation boundary can be chosen without adding a corresponding ID to requests:

The example in Section 3 bounds a tenant's total allocation while the origin checks each Credential separately. Applying that policy to a project requires the Anchor to account for all allocations to that project.

The Anchor can also cap the sum of Endorsements across projects belonging to one product. The origin enforces each Credential's ten-credit limit without seeing the project or product. The project and product totals depend on the Anchor's issuance policy, not on an origin-side counter keyed by an ID. Additional accepted Anchors or Endorsements increase the available budget unless their policies constrain the combined allocation. Public-baseline requests and the platform's aggregate limit remain separate from this bound.

These budgets replace identifiers only for the capacity decisions their issuance and spending policies implement. An origin cannot use an anonymous presentation to identify and block a particular project, audit its activity, or correlate all requests in a session. If it needs those functions, signed metadata can accompany MOLE, making requests linkable at the metadata's scope.

Authors' Addresses

Thibault Meunier
Cloudflare
Christopher A. Wood
Cloudflare