When an AI agent hands a task to another agent, which then calls a service, which calls another, nobody in the chain can easily say that these are all parts of one conversation. A new IETF working group, Agent Communication Protocols (agentproto), exists to fix that, and its charter is approved and the group listed as active, with the page last updated on 8 October 2026. It has a charter and no protocol yet. Watch it, and do not design around it.
What it is
The charter defines a dialog as "the correlated sequence of interactions between two or more participants over the course of a task". The group is to write an Agentic Dialog Management Protocol, on the IETF's Standards Track, that covers four things: identifiers that tie related interactions together across several hops, trust boundaries and time; a lifecycle for starting, changing, ending and revoking the sharing of that context; recovery when a network path or an intermediary fails or a participant cannot be reached; and bindings, meaning how the dialog context rides on existing protocols such as HTTP, QUIC, WebTransport, WebRTC and MoQ (media over QUIC). The responsible Area Director, the IETF official who shepherds the group, is Charles Eckel.
What it does not do
The charter says it does not replace application-layer agent protocols or define a new transport. Protocols such as MCP and A2A can carry the dialog context in their own way, and any change to them is for their maintainers. Also out of scope: how agents are built, models, planning and decision-making, preventing a model's bad behaviour, user interface design, and the format of audit records (though dialog identifiers may be used in auditing). Two informational documents will accompany the protocol: a reference architecture and a use-cases and requirements document.
Where it falls short, so far
Two things limit how much to read into this. The charter page lists no milestones or dates, so there is no schedule to plan around. And an identifier that follows a dialog across organisations can also be used to link activity, which the charter itself flags: the group is to analyse those privacy risks and specify mitigations. I did not check whether the group has published any draft yet.
What to do
Watch. A standard way to carry one dialog's identity across systems would help when an agent hands work on, or resumes after a break. Follow the group's mailing list if you build multi-agent systems; do not design around it until it publishes a draft.