Review · November 22, 2026
"PACE: Protocol for Agent Communication Exchange (MOQT)"
Read This First
This draft is no longer active. Datatracker marks it as replaced by draft-nandakumar-agentproto-pace, now at revision 01.
This review covers the June 2026 revision as the starting point of PACE. Anyone preparing feedback should read the successor.
The Problem
Agent protocols like A2A and MCP each bring their own transport, session, and identity story. None of them handle real-time audio and video alongside tool calls. None of them fan out to many agents through shared infrastructure.
PACE proposes one session layer for all of them, built on Media over QUIC Transport (MOQT). Agents connect to a relay, publish on their own tracks, and subscribe to each other's. A2A, MCP, and AutoGen messages ride inside PACE sessions unchanged.
The draft also takes on agent discovery, agent identity tokens, and delegated authorization.
How It Fits In
This was submitted to AGENTPROTO. Part of it matches the charter and part of it runs well past it.
The session lifecycle and the MOQT binding are in scope. The charter asks for dialog lifecycle primitives and names MoQ as a candidate transport binding.
Three things are outside it:
- Identity and authorization. The charter says the reference architecture should reuse existing identity, authentication, and authorization building blocks "rather than defining new ones". This draft defines a new token format, a new scope syntax, and a new delegation model.
- Discovery. Agent discovery belongs to coordination with other groups, not to the dialog protocol.
- Application protocol bindings. The charter leaves changes to A2A and MCP to their maintainers. This draft maps every A2A operation onto MOQT and adds a
moqtblock to the A2A Agent Card.
The dependency list is short and mostly unfinished:
- draft-ietf-moq-transport, the only normative protocol dependency besides QUIC and TLS
- draft-ietf-moq-c4m for relay authorization
- The authors' own MCP over MOQT and SD-JWT agent identity drafts
- A2A 0.3.0, MCP 2025-06-18, and AutoGen 0.4
So how does it relate to OAuth? It doesn't. The only OAuth-family reference is JWT (RFC 7519). The string "oauth2" appears once, inside an example Agent Card. There is no authorization server, no Token Exchange, and no DPoP.
What It Actually Does
Everything is a MOQT namespace under ("pace", "agents", {agent-id}).
Discovery. Each agent announces its namespace and publishes an Agent Card on a card track. Any agent can subscribe to the ("pace", "agents") prefix and learn about every agent on the relay network. An agent then "follows" a peer by subscribing to everything under that peer's prefix.
Sessions. A session is a UUID that any agent can generate. Establishing one works like this:
- Agent A publishes
SESSION_INITon actrltrack under its own namespace plus the session ID. The message names the target and carries A's identity token. - The relay forwards it to B, because B already follows A.
- B publishes
SESSION_ACCEPTon its ownctrltrack under the same session ID. - Both sides publish further tracks:
chat,audio,video, and oneops/{req-id}track per request.
Each agent only ever publishes under its own namespace. A response is a track with the same name under the responder's namespace.
Bindings. Application protocols get a track name prefix: a2a/, mcp/, autogen/, or ext/{proto}/. Messages keep their native JSON-RPC format inside a small PACE envelope.
Quality of service. Eight priority tiers are carved out of MOQT's 0-255 range. Session control comes first, then media, then chat, then tool calls, then bulk transfer.
Identity and authorization. Each agent presents a JWT with an agent ID, issuer, subject, capabilities, scopes, and a delegation-chain. Scopes take the form {domain}:{resource}:{action}:{constraints}. Each delegation hop narrows the scope. High-risk operations can require "cryptographic evidence of approval".
What Should OAuth People Look At?
The authorization model reinvents work that already exists, and it does so in three paragraphs.
- The token claims ignore the JWT registry. The draft says the token is "based on JWT" and then names the claims
issuer,subject,issued-at, andexpires-at. No JWT library will validate those. delegation-chainduplicates theactclaim. Token Exchange already defines nested actor claims for exactly this path.- Nobody issues the narrowed token. "Agent A requests a sub-scope for Agent B." From whom? No issuer role, endpoint, or exchange is defined.
- The scope syntax competes with Rich Authorization Requests. The four-part string is given one sentence and no grammar.
- The token has no sender constraint. It is a bearer credential.
That last point becomes a real attack once you look at where the token travels. SESSION_INIT carries the identity token on A's ctrl track. The relay forwards that track to every agent that follows A, not only the target. The relay also caches it for late joiners. Any follower can lift the token and present it in its own SESSION_INIT.
Privacy has the same shape. The session discovery namespace publishes each session's description and full participant list to anyone who subscribes.
The Security Considerations section is one word: "TODO".
Interoperability Concerns
No session message has a wire format. SESSION_INIT is a bullet list of field names with no types and no encoding. The other eight message types have names and code points only. Two implementations could not complete a session handshake.
The message set is inconsistent. The IANA table registers nine session messages. The ctrl track description lists four. SESSION_REJECT, SESSION_SUSPEND, and SESSION_RESUME appear in the state machine or registry and nowhere else. SESSION_UPDATE requires an acknowledgment from all participants, and there is no acknowledgment message.
The envelope flags do nothing yet. Compressed, fragmented, and encrypted are each a bit. No compression algorithm, fragmentation scheme, or encryption scheme is named.
Agent IDs are unanchored. The text says UUID. Every example uses names like agent-a and fs-server. Nothing ties the namespace an agent publishes under to the agent ID in its token. The draft says relays enforce this with C4M tokens and specifies none of it.
Requests have no addressee. A request is published on ops/{req-id} and reaches every subscriber. In a three-agent session, nothing says which agent should answer.
The MOQT reference doesn't match the mechanism. The draft cites draft-ietf-moq-transport-16. The whole design depends on SUBSCRIBE_TRACKS, which is not in revision 16 and appears only in later revisions.
The MCP binding points at a moving target. It defers to MCP over MOQT, then remaps that draft's namespace tuples into track names. Any change there breaks the mapping here.
Recovery relies on undefined pieces. The failure table uses a session heartbeat that is never defined. Checkpoints go on "a dedicated track" with no name or format.
What Could Be Removed
The identity token and authorization model. Reference existing work and state what PACE needs from it. This removes the charter conflict as well.
The AutoGen binding. AutoGen is a framework, not a wire protocol. The message format shown is invented for this draft.
The A2A and MCP bindings. Keep the binding architecture and the registry. Move each mapping to its own document, worked out with that protocol's maintainers.
Global agent discovery. Enumerating every agent on a relay network is a discovery design with its own privacy problems. Sessions work fine with an agent ID learned out of band, which the draft already allows.
The "Why MOQT?" list, the scalability section, and the failure table. These argue for the design instead of specifying it.
The priority tables. One table of tiers is enough. The per-binding tables repeat it.
What remains is a session lifecycle and a namespace convention over MOQT. That is the part the charter asks for.
Writing Quality
The prose is clear and reads easily. The namespace design is explained well, and the appendix walkthroughs make the model concrete.
It is an outline more than a specification. Most sections are a paragraph or a table. BCP 14 keywords are rare, and several land in odd places, such as a "MAY" inside the list of reasons to use MOQT.
There are signs of a fast edit. The "latest revision" link points to a different draft name. The MCP over MOQT reference is to draft-mcp-over-moqt, which is not a real draft name. RFC 9297 is cited for WebTransport and is the HTTP Datagrams spec. RFC 7540 and RFC 9110 are listed as references and never cited. Acknowledgments is also "TODO".
At 36 pages the length is fine. The problem is the spread: ten topics at outline depth, where two at full depth would be reviewable.
Authors
Both authors work at Cisco and come from the real-time media side of the IETF. Neither has published an RFC in OAuth.
Suhas Nandakumar has 9 published RFCs from 2013 to 2025, all in real-time media: STUN and TURN URI schemes (RFC 7064, RFC 7065), RTP taxonomy (RFC 7656), SDP and ICE (RFC 8839, RFC 8859). There are 36 active Internet-Drafts on the profile. About two thirds are Media over QUIC, including the transport draft. Around ten are recent AGENTPROTO, OAuth, and WIMSE drafts. Nandakumar is writing at very high volume right now.
Cullen Fluffy Jennings has 29 published RFCs from 2002 to 2026, including SIP identity (RFC 3325, RFC 4474), NAT behavior (RFC 4787), MSRP (RFC 4975), and JSEP for WebRTC (RFC 8829). There are 30 active Internet-Drafts on the profile. Half are Media over QUIC. Eleven are agent, OAuth, or WIMSE drafts, most co-authored with Nandakumar.
Bottom Line
The core idea is sound. Per-agent namespaces on a pub/sub relay give multi-party sessions, media, and tool calls one model, and the authors know MOQT well.
This revision is a sketch. No session message has a format, and the security section is empty.
The identity and authorization sections are the part to push back on. They define a parallel token and delegation model with no reference to OAuth, and the token travels as a bearer credential on a track that relays fan out and cache.
The good news is that the authors appear to have reached the same conclusion. The abstract of the replacement, draft-nandakumar-agentproto-pace, drops identity from PACE's scope and moves delegation to a companion document. Review effort belongs there.