2026-07-04: -00 → -01 bump (July 3) — the follow-on revision the first-look flagged to watch for.
draft-liu-moq-live-agent-interaction-01was posted July 3 2026, five days after the -00 submission. It remains an individual draft (not adopted; RFC stream “None”, IESG state “I-D Exists”), and the Datatracker shows no explicit -00→-01 changelog; the abstract is unchanged in shape (ASR-transcript / LLM-token / TTS-audio mapping onto MOQT’s Group/Subgroup/Object hierarchy, turn-taking with barge-in, a pure application profile). Notable mainly as continued Alibaba investment in the “MoQ as a reusable real-time substrate” thesis — a second revision inside a week — but still undiscussed on the moq@ietf.org list. Logged as a revision event; watch for any call for adoption or WG-list reaction. See discussions-2026-07.2026-06-30: First-look — a new individual I-D applies MoQ to real-time AI-agent (voice) interaction, the first MoQ + AI-agent draft the wiki has tracked.
draft-liu-moq-live-agent-interaction-00“Live Agent Interaction over MoQ” was submitted June 29 2026 by Yanmei Liu (Alibaba Inc.) and Dapeng Liu (Alibaba Cloud) — the second Alibaba MoQ artifact alongside the XQUIC implementation. It is a pure application profile: it adds no transport-layer mechanism, instead defining semantic conventions for mapping a live conversational AI session onto MOQT’s object hierarchy. Brand new and not yet discussed on the WG list; logged here as a first-look pending any uptake.
draft-liu-moq-live-agent-interaction-01 | individual submission | -00 submitted 2026-06-29, -01 posted 2026-07-03
Authors
- Yanmei Liu (Alibaba Inc.)
- Dapeng Liu (Alibaba Cloud)
Abstract (summary)
Defines an application-layer protocol for real-time interactive communication between users and AI agents over Media over QUIC Transport. It specifies how streaming inference outputs — ASR transcripts, LLM tokens, and TTS audio — map onto MOQT’s object hierarchy, defines turn-taking control with interruption (barge-in) support for voice conversations, and establishes track-structure conventions for live agent sessions. It operates as a pure application profile: it does not modify underlying transport semantics.
Key Ideas
- Maps the conversational structure onto the MOQT data model:
- conversational turns → Groups
- inference steps → Subgroups
- token batches → Objects
- Leverages MOQT primitives directly: publish/subscribe, prioritized delivery, group-based organization, and the relay infrastructure.
- Turn-taking + barge-in: turn-control mechanisms with interruption support, the latency-critical requirement for natural voice conversation.
- Relay-transparent: because the structure is expressed in MOQT’s native Group/Subgroup/Object hierarchy, relays can route AI-agent traffic correctly without payload inspection — no new relay behavior required.
Why it matters
- First MoQ + AI-agent draft in the WG’s document space — extends MoQ’s framing beyond live media (video/audio/data) into conversational AI / voice-agent transport, a distinct application class.
- Demonstrates the “MoQ as a reusable real-time substrate” thesis: rather than a new wire protocol, it is a convention layer on top of the existing transport — the same “pure application profile” stance the WG has favored for keeping the core spec lean as it approaches WGLC.
- A second concrete Alibaba contribution to MoQ (after xquic-moq).
Status & Caveats
- Individual draft, now at -01 (posted July 3 2026; -00 June 29 2026) — not adopted, not yet discussed on the moq@ietf.org list as of this writing; no explicit -00→-01 changelog on the Datatracker.
- The wiki tracks individual drafts that are actively discussed or referenced; this one is logged as a first-look because it opens a new application category for MoQ. Watch for any WG-list reaction or follow-on revision.
Related
- moq-transport — the transport this profile runs over (turns→Groups, steps→Subgroups, tokens→Objects)
- moq-msf / moq-cmsf — streaming-format layers for media payloads (this draft is an application profile, not a media container)
- xquic-moq — Alibaba’s MoQ implementation
- subgroups-and-objects — the Group/Subgroup/Object data model this draft maps onto