Field guide to the AI Agent Protocol CG
The group was founded to write open protocols that let AI agents discover, identify and collaborate with each other across the Web. Its starting material is the Agent Network Protocol (ANP), an open-source design by co-chair Gaowei Chang. Song Xu of China Mobile is the other co-chair. For most of its life the list carried a handful of messages a month, mostly meeting notices and minutes.
October 2026 brought 96 messages. Most of them come from a small set of people who each maintain their own draft specification, published elsewhere (IETF Internet-Drafts, GitHub repositories, other W3C groups). They use the list to compare designs, and each brings the vocabulary of his own draft. Terms like lease, standing read, envelope or sidecar are private coinages from one draft, used without definition because the author assumes readers have opened the link.
Three older vocabularies run underneath. W3C identity work supplies DID and verifiable credential. Web security supplies OAuth, client_id and certificate checking (OCSP). Software testing supplies test vectors, digests and conformance suites. Once you can tell which vocabulary a sentence is speaking, most threads become legible.
The September proposal on governance gaps opens by citing a "Layer Zero" discussion in the AI Knowledge Representation CG. Some of this traffic therefore crosses over from groups you already know. For a reference on the identity and trust layers, see the AIKR CG draft technical note The Knowledge Representation of the Trust Layer Through the Lens of Digital Identity Chains, open for consultation.
Map of the stack
Each layer answers one question an agent faces when it deals with another agent. Threads can be placed on this map; October's long threads all sit in the shaded rows.
An agent needs a name that it controls and can prove. The group uses W3C Decentralized Identifiers. ANP defines its own DID method, did:wba (Web-Based Agent). It runs on ordinary web servers. ANP also adds human-readable WNS handles that resolve to a DID.
An agent publishes a structured description of its functions and interfaces, and others find it through a discovery mechanism. July minutes record the group's recommendation to separate discovery from flows that assume the remote agent's URL is already known.
ANP adds a meta-protocol layer in which agents negotiate, in natural language, which protocol to use, plus end-to-end encrypted messaging. Outside the group, MCP connects an agent to its tools and A2A connects agents to each other; both now sit under the Linux Foundation's Agentic AI Foundation.
Proving who you are does not grant permission. An ANP issue spells this out: did:wba authentication identifies the caller, and approval is left to higher layers. The group's open proposal (PR #46) separates roles and direct mandates from ordinary OAuth access permission.
Open decision in the group
Permission can expire or be withdrawn. The October threads argue about how long a verified statement may be trusted, what happens when revocation lands while an agent is mid-task, and what to record when the revocation service cannot be reached.
Most of October's traffic
Audit trails, decision receipts, provenance records and legal accountability. Some contributors want enforcement at the moment of action; others want a verifiable record for later review. A regulatory practitioner argues that identity alone says nothing about who is liable.
Lease experiment
The longest thread turned into a shared test. Larry Lewis (chair of the Agent Trust Protocol CG) proposed one timeline that every participant runs against their own design. Play with it below.
The scenario: at second 0 an agent receives valid identity and authorization evidence. At second 10 the issuer revokes it. At seconds 30, 60 and 119 the agent tries to act. The numbers come from ShaoBao Li's draft, where every statement carries a 120-second freshness lease. Tymofii Pidlisnyi (Agent Passport System) pointed out the weak spot: an action at second 119 can rely on a statement its issuer revoked at second 10.
Li's reply describes two modes with different trade-offs. A cached lease keeps working offline and caps the exposure at one window. A live read refuses as soon as revocation is visible, and refuses again when the service is silent, because an unreachable authority is not read as consent. The participants then agreed to record results along separate axes so that each design is scored only on what it claims to do.
Refuse or halt
Larry Lewis's framing has become the reference point for the whole discussion. He asks implementers to keep three questions apart.
| Question | What it asks |
|---|---|
| Evidence freshness | How recently was the statement or credential verified? |
| Authority validity | Is the delegation still valid and unrevoked? |
| Action-time authorization | Is this specific operation permitted now, under the constraints that apply? |
Li's draft claims the first two and leaves the third to the "calling layer", the software that actually performs the action. When it refuses, it must give one of five reasons, each a fact a third party can recompute.
| Refusal reason | Meaning |
|---|---|
| Evidence did not match | The presented proof fails to check. |
evidence_expired | The lease ran out. |
| Authority not determined | The verifier could not reach the source. Kept separate so a record never invents a decision nobody made. |
authority_revoked | The grant was withdrawn. |
| Operation not permitted | The grant exists but does not cover this act. |
Steven Zhao (Zhao Xinghua) then asked about the agent already halfway through a job when revocation lands. A refusal stops the next call at the boundary, he argued, but a running chain makes no new call and never meets that boundary. It needs a halt transition, demonstrated in his runnable state machine:
The chain pauses at the next safe point between sub-calls, rolls back to its last checkpoint, and resumes only when a registered restore condition is met. Without one, he warns, a temporary problem kills the job for good, as with eIDAS Article 28(4), under which a revoked certificate can never be reinstated. He borrowed the failure vocabulary from OCSP, the long-standing certificate status protocol, where "unknown" is distinct from "good" and tryLater marks a responder that exists but cannot answer yet.
The open question at month's end: where exactly is the safe interruption point, and should the group define it or leave it to each runtime?
Thread decoder
October's threads, the person who started each, and what they are arguing about in plain terms.
A review of the group's Agentic Web report with four gaps: detecting a swapped identity document (preimage defense), who vouches for the address of a description (address assurance), whether trust expires (trust leases), and whether interoperability works platform by platform or statement by statement (granularity). It became the joint revocation test described above, with Agent Passport System, Agent Trust Protocol and Agent Authorization Envelope authors taking part.
Should the protocol define what counts as trustworthy at all? Ren argues for reference methods agents may choose, since a hard-coded check becomes a hole once it is broken. Alan Karp replies with prior questions: does the agent decide, or its human operator, and what happens when agents share credentials?
What a running agent must do, and what its receipt must record, when authority disappears or cannot be checked mid-task.
NTI is an enforcement layer that intercepts an agent's tool calls and blocks unsigned or out-of-scope ones. Alan Karp points to HDP, an IETF draft that keeps an audit trail. Abisheak quotes HDP saying its tokens must not drive access decisions, and calls the two complementary: one blocks at the door, the other records who entered. NTI also publishes a daily leaderboard scoring agent frameworks against NTI's own standard.
A request for someone attending the Agentic Internet Workshop in Mountain View to raise two topics: delegation chains that cross MCP and A2A, and keeping authority revocable and auditable when every hop looks compliant but the whole chain exceeds its mandate.
The group's own business. Gaowei presented the Agent Authorization proposal (PR #46). Two designs are open: use a DID as the OAuth client_id, or keep CIMD's HTTPS client_id and bind a DID to it. No decision was recorded. Song reviewed 15 TPAC 2026 Agentic Web breakout proposals; a charter for a proposed Agentic Web Working Group is still being drafted.
Cost and budget fields so a calling agent can cap what a delegated task spends (Mehek Mandal). Heartbeats that suspend an agent's authority when it goes silent (Athar Nouman, Living Agents Protocol). An open map of agentic-commerce standards across W3C, IETF, GS1 and industry (Francesco Marinoni Moretto). A TPAC breakout on browsers for spatial computing and robotics.
Glossary
Each entry links to the message or specification it is drawn from.
Reading order
- Read the meeting minutes first. Gaowei Chang posts a summary after each biweekly call, and these record what the group itself has discussed and decided. Proposals in other threads carry no group endorsement until they appear here.
- Keep the protocol draft and the ANP documentation open. They define the identity, description and discovery layers the group actually owns.
- For any long thread, find the first message and the draft it links. Each participant writes in the terms of his own draft, so the draft is the dictionary.
- Place each message on the stack map. Ask which of the three questions (freshness, validity, action-time authorization) it addresses and which it leaves to someone else.
- Skip calendar notices ("Event Updated") unless you plan to attend.