The Scope Creep Automated reviews of new AI-related working group submissions

Review · November 22, 2026

"Requirements for Agent Session Establishment, Capability Negotiation, and Sessionless Interaction"

draft-feng-agentproto-session-requirements -02View on Datatracker →

The Problem

Discovery work like DAWN stops once two entities have found each other. This draft tries to define what happens next.

It is a requirements document. It says what a "session" between two agents must be, how it gets established, and what rules apply to one-off messages that don't use a session. It defines no messages, no formats and no protocol.

The core idea is that a session is not a connection. A session is an association between two authenticated peers, with a negotiated capability set and authorization state that changes over time.

How It Fits In

Does this fit the agentproto charter? Partly, and the part that doesn't fit is the main design choice.

The charter covers dialog lifecycle, transport bindings, and a use cases and requirements document. The draft offers itself as input to that requirements document. Its transport independence, establishment, failure handling and resumption requirements line up with the charter's dialog lifecycle work.

The problem is the unit of work. The charter defines a dialog as interactions "between two or more participants" and wants context carried "across multiple intermediaries" and trust boundaries. This draft defines a session as strictly bilateral and puts delegation, handoff and third parties out of scope.

The working group's adopted use cases and requirements draft says the gap is exactly the opposite. MCP sessions and A2A contexts are already scoped to a single hop. The missing piece is correlating a task across hops. This draft specifies the single-hop piece in detail and leaves the multi-hop piece to someone else.

The draft never uses the word "dialog". It doesn't say whether a session is a dialog, part of a dialog, or something underneath one.

The charter also says identity, authentication and authorization should be reused from existing work, not defined new. The adopted requirements draft hands per-action authorization to the OAuth working group. This draft requires an authentication negotiation phase, a capability negotiation phase and an in-session authorization exchange. All three imply new protocol machinery inside agentproto.

There are no dependencies on any RFC beyond BCP 14. OAuth, WIMSE, MCP and A2A are not cited. The informative references are two DAWN drafts, draft-schrock-ae-challenge, and two of the author's own drafts on intent routing and the Agentic Intent Network.

Summary

There are 11 requirements covering two interaction modes.

Session-based interactions go through fixed phases:

  1. Authentication. Each side declares what authentication it needs and supports. They pick something mutually acceptable and authenticate each other.
  2. Capability negotiation. Each side declares capabilities and requirements. Both must accept the resulting set.
  3. Parameter agreement. Everything so far is bound to one prospective session.
  4. Establishment. The session exists only when both sides have accepted and both can tell it completed.

An established session starts with "minimum session authority". That lets a peer ask for authorization and nothing else. Either peer can then request authorization from the other for an operation, and the other can grant, deny or constrain it. Grants can be revoked or replaced at any time.

Roles are per request. Whoever opened the session is not permanently the client.

Capabilities can be added through renegotiation without a new session. A lost capability stops being usable immediately. A session can survive a change of connection or runtime instance, but only for the same two authenticated entities. A different entity always needs a new session.

Sessionless interactions are notifications, probes and atomic requests. Each must carry its own security context and can't inherit anything from a session, even one on the same connection.

So how does this relate to OAuth? In OAuth terms, a sessionless request with a side effect is an ordinary API call with an access token. Minimum session authority looks like a token with no scopes. Requesting authorization per operation looks like Step-Up Authentication combined with Rich Authorization Requests.

The difference is who decides. In this model the peer is the authority, and a user or "external authority" is an optional input. There is no authorization server role. A deployment where an AS issues the token before the session starts isn't ruled out, but the draft doesn't say how it maps.

Interoperability

Significant interoperability concerns

There is nothing to implement. That is by design for a requirements document. The useful question is whether a protocol designer could test a proposal against it. Mostly they can't.

The requirements have no consistent target. Some bind "the interaction model". Some bind "a transport binding". Some bind entities at runtime, such as "an entity MUST NOT treat expired or invalidated information as current". There are around 80 normative statements across these three kinds of target.

Many requirements are not testable. Cached information needs freshness semantics "appropriate to the information being cached". Integrity and confidentiality are required "when required by the risks of the interaction". The effect of a capability change "MUST be explicit". None of these let a reviewer say a protocol passes or fails.

There are no use cases. The adopted working group draft ties each requirement to a scenario. This draft has none. Mandatory capability negotiation, per-request role switching and peer-issued authorization are asserted without showing who needs them.

The phases don't match deployed protocols. A2A has no session establishment or capability negotiation. An Agent Card is fetched out of band, and under this draft every A2A call would be sessionless. MCP negotiates capabilities in initialize, but authorization happens at the HTTP layer through OAuth, not as a negotiated phase between the peers. The draft doesn't say whether these count as conforming.

Authentication negotiation has no downgrade requirement. The security considerations name the attack: an attacker removes a strong method or inserts a weak one. The only defense is each side's local policy. The negotiation happens before authentication, and nothing requires it to be verified afterward.

Capability negotiation is underspecified where it matters. The draft requires a "negotiated capability set" that both sides can identify, but explicitly requires no format. Two protocols could both satisfy REQ-3 and share nothing.

Mechanisms that could be removed

  • REQ-4 and REQ-5. Both say each side makes its own decisions and can decline. They could be one short requirement.
  • REQ-10 on caching. What an entity caches and when it revalidates is local behavior. The peer can't observe it.
  • The mandatory authentication negotiation phase. Most deployments learn the required authentication from metadata before connecting, the way OAuth clients read authorization server metadata. Making it a session phase forces a round trip that discovery already covers.
  • CDL and the AIN reference. The draft says CDL is not required. It only serves to point at the author's other work.
  • The out-of-scope material. Delegation, handoff and audit are declared out of scope in the abstract, introduction, scope section, REQ-9, the security considerations and the change log. One paragraph would do.
  • The "determinability or audit profile" paragraphs. They describe what some other document might require.

The strongest parts are REQ-1, REQ-6 and REQ-7. A session is not a connection, state from one session can't leak into another, and both sides must agree on whether a session exists. Those are real failure modes in multiplexed transports. They would make a good short contribution to the adopted requirements draft.

Writing Quality

The prose is formal, grammatical and reads like IETF text. Each requirement has a rationale paragraph, which helps.

It is repetitive. "MUST NOT by itself" and "MUST NOT silently" appear throughout, and the bilateral session point is made at least six times. The 22 pages could be 10.

Some specific issues:

  • A normative MUST about what a human is shown appears only in the Security Considerations, not in any numbered requirement.
  • The Entity definition uses "MAY" to list what an entity can be.
  • An entity can be a human. The draft doesn't explain how a human takes part in capability negotiation.
  • References have no authors and no URLs. The DAWN terminology and problem statement references point at -04, and both have newer revisions.
  • The page header title is "Agent Interaction Requirements", which doesn't match the document title.
  • The change log covers only -01 to -02.

The draft went through three revisions in five days in August, then stopped. It has not been updated since the working group adopted its own requirements document.

Author

Chong Feng is the sole author, with no affiliation listed.

  • One published RFC: RFC 10016, "System-Defined Configuration" (September 2026).
  • Nine active Internet-Drafts, which is a lot of concurrent writing.
  • Most of that work is in network management: NETMOD, NETCONF, NMOP and NMRG.
  • Two other drafts target agent protocols: draft-feng-agentproto-idl-info-model and draft-feng-dmsc-intent-routing-requirements.

The author has IETF experience, mainly in YANG and network operations. The agent drafts form a connected set: intent routing gets a request to an entity, this draft covers the session that follows, and IDL describes the agent. This one is the most protocol-neutral of the set, and its transport and state-binding requirements are the pieces most likely to be picked up by the working group.