Review · November 22, 2026
Live Agent Interaction over MoQ
The Problem
Voice agents need more than a token stream. The user talks continuously, the agent answers with text, speech, and tool results at once, and the user can interrupt mid-sentence.
The draft argues that SSE and WebSockets cannot do this well over TCP. Media over QUIC Transport (MOQT) has the right primitives, but no conventions for turns, interruptions, or inference output. This draft defines those conventions as an application profile on MOQT.
How It Fits In
This is not an OAuth draft. It is an individual submission aimed at the MoQ working group, and it sits with the AGENTPROTO-related work.
It replaces the approach of Dapeng Liu's earlier Agent Protocol over MoQ, which wrapped its own frame format around MoQ. This draft maps directly onto MOQT's tracks, groups, subgroups, and objects, and cites the older draft only informatively. That puts it in line with MCP over MOQT, which takes the same approach for MCP traffic.
The dependencies are all MoQ working group drafts: MOQT and LOC normatively, MSF and Secure Objects informatively. It cites no OAuth RFCs and no token format.
Revision -02 adds a per-action authorization exchange for tool calls. That is the part OAuth people should look at.
What It Actually Does
There are two roles, a user device and an agent backend, with an optional MOQT relay between them.
A session is a namespace with a fixed set of tracks:
input/audioandinput/textfrom the useroutput/audio,output/text, andoutput/toolfrom the agentcontrol/userandcontrol/agent, one control track in each direction
The central mapping is that one conversation turn is one MOQT Group. A sentence or tool call within the turn is a Subgroup. A batch of tokens or an audio frame is an Object. The same Group ID is used across tracks for the same turn, so a relay under congestion drops old turns before new ones.
Each track gets a small binary payload format. Text objects carry a flag (partial, final, cancelled) and a token delta. Audio objects wrap a LOC payload with optional fields pointing at the text being spoken. Tool objects carry a status, numeric tool and call IDs, and JSON.
Turn-taking runs on the control tracks. There are nine signals, including SPEECH_START, SPEECH_END, TURN_STARTED, and TURN_COMPLETE.
Interruption is the best-specified part. The client sends BARGE_IN as a datagram at top priority and mirrors it on the reliable control track. The agent deduplicates by event ID, stops output, marks the turn cancelled, and replies INTERRUPT_ACK with the position where it stopped. The draft also handles the race between BARGE_IN and TURN_COMPLETE.
So What Does Authorization Look Like?
There are two layers, and neither is defined.
Session access. Endpoints "MUST authenticate both the user and agent endpoints". How is not stated. MOQT's authorization token parameter SHOULD be used for per-track authorization. The draft says nothing about the token format, the issuer, or how a token maps to track names. The MoQ working group already has C4M for this, and the draft does not cite it.
Per-action authorization. When a tool call needs approval, such as a payment, the agent sends AUTH_REQUIRED on its control track. The user side answers with AUTH_RECEIPT. Both carry a tool ID, a call ID, and an opaque "artifact".
The artifact can be "an offline-verifiable receipt, a challenge reference, or another application-defined authorization artifact". That is the whole definition. The draft gives the envelope and leaves the contents to someone else.
Liu has several individual drafts in OAuth on per-action authorization and evidence. None is cited here, so the reader cannot tell what is meant to go in the field.
The surrounding rules create real exposure:
- Anyone on
control/usercan approve. The receipt comes from "the user endpoint or authorizing component". Nothing says who may publish on that track or how the agent verifies the receipt. - Relays read the approval traffic. The draft says control tracks SHOULD NOT be end-to-end encrypted. Challenges and receipts for payments pass through every relay in the clear.
- Observers see tool results. The draft promotes multi-subscriber fan-out for monitoring and recording. Nothing limits who can subscribe to
output/toolorcontrol/agent. - User audio has no end-to-end option. The draft calls it sensitive biometric data, but only offers Secure Objects for agent output.
- Approvals can be missed. Relays SHOULD NOT cache control objects. A client that reconnects during a pending approval never sees
AUTH_REQUIRED, and nothing defines a timeout or a denial signal.
Session IDs MUST be cryptographically random, with UUIDv7 recommended. UUIDv7 leads with a timestamp, and the ID sits in the namespace where every relay sees it.
Interoperability Concerns
The text and barge-in formats are close to implementable. The rest is not.
Session setup is missing. The draft never says who picks the session ID, how the agent learns which namespace to subscribe to, or how a session ends.
The namespace is written as a URL. moqt://{authority}/agent/{session-id}/ is given as the Track Namespace. MOQT namespaces are tuples of fields, and the draft does not say how the URL maps to one.
The control track has no object mapping. Turn IDs travel in the payload. The Group and Object IDs of control objects are never defined, including for the BARGE_IN datagram.
Signal payloads are undefined. INTERRUPT_ACK is shown with three named fields in a diagram and no encoding. TURN_STARTED and THINKING have no payload definition. The wall-clock timestamp has no stated use.
Turn ID allocation is unclear. User audio starts a new Group at every voice-activity trigger. A pause mid-sentence or a false start burns a Group ID, and the draft does not say which one the agent's reply uses. Section 2.2.4 refers to a "turn counter" that Section 3.1 never defines.
Tool calls are numbers with no meaning. tool_id is a varint. No track or catalog entry maps it to a tool. The JSON metadata and payload have no schema, and the "authorization reference" that metadata MUST include has no field name.
Timeouts are promised and absent. Section 5 is titled Delivery Policies and says it specifies per-track timeouts. It contains one table on datagrams versus streams and no timeout values.
Relay transparency contradicts itself. The draft says no relay modification or payload inspection is needed. It also says relays MAY perform voice activity detection, and that control tracks stay unencrypted because "relay-level inspection may be needed for priority enforcement".
Smaller conflicts. Client VAD signals SHOULD go by datagram in Section 3.4, and by stream with no fallback in Table 5. Two track properties are registered with IANA and never used in the body. input/text and the video tracks are named with no format. The abstract promises a mapping for ASR transcripts, and Section 2.2.3 defers them to a future extension.
What Could Be Removed
The audio envelope. Wrapping LOC in a new container means a LOC-only receiver cannot play the audio, which the draft admits. The alignment fields could be an extension inside LOC or a separate timing track.
seq and count in text objects. MOQT Object IDs already order objects in a subgroup. Nothing uses the token count.
The opaque authorization exchange. An envelope with undefined contents does not interoperate. Either define the artifact by reference to a real spec, or move this to a separate draft and keep the transport profile small.
Batching intervals, priority tables, and the 50ms processing target. These are implementation advice. They belong in an appendix without normative keywords.
THINKING, the padding suggestion, and the two unused track properties.
Design principles and Appendix B. They restate each other and the introduction.
Writing Quality
The English is clean and the structure follows IETF conventions. BCP 14 keywords are used throughout, and the diagrams and examples are helpful.
It is too long for what it specifies. The Group, Subgroup, and Object mapping is presented as a table twice. The document outline appears twice. The introduction, design principles, and rationale appendix cover the same ground three times. At 34 pages, the wire formats take up about five.
Several normative statements cannot be tested. "The agent MUST process BARGE_IN within one processing cycle" and "the agent MUST keep the authorization artifact transport-agnostic" are examples. Words like "USES" and "DOES NOT" are capitalized as if they were keywords.
The barge-in section shows what the rest could be. It has a payload format, deduplication rules, and a defined answer for the race condition. The new authorization section was added without that level of care.
Authors
Yanmei Liu (Alibaba) has 1 published RFC, RFC 10069 on Multipath QUIC, published this month. Liu has 4 active drafts and 6 expired, all in QUIC and MoQ. This is a transport person with a recent RFC, not a high-volume author.
Dapeng Liu (Alibaba Cloud) has 3 published RFCs from 2014 and 2015, all on mobility management. Liu chairs the DAWN working group and has 77 expired drafts. There are 14 active drafts, most of them OAuth and AGENTPROTO individual drafts on agent authorization. Liu is writing at high volume right now.
Neither author has published an RFC in OAuth or MoQ.
Bottom Line
The core idea is sound. One turn per Group is a natural fit for MOQT, and the barge-in design is careful work. This is a clear improvement on Liu's earlier frame-based draft.
It is not yet implementable. Session setup, the control track mapping, tool identifiers, and delivery timeouts all need to be written down, and the document could lose a third of its length.
For OAuth people, the thing to watch is the new AUTH_REQUIRED and AUTH_RECEIPT pair. It is a transport-level slot for per-action user approval with nothing defined inside it and no protection around it. If per-action authorization for agents gets standardized anywhere, this draft shows where it would be carried over MoQ, and the receipt would be crossing relays unencrypted on a track with no publisher rules.