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

Review · November 22, 2026

"IDL: A Semantic Information Model for Agent Communication"

draft-feng-agentproto-idl-info-model -00View on Datatracker →

The Problem

OpenAPI, MCP tool schemas and A2A Agent Cards describe what an endpoint does when you call it. This draft argues that an agent is not an endpoint. An agent can start interactions, negotiate, refuse, and escalate to a human.

IDL (Intent Description Language) is a proposed metadata model for describing those properties. It answers "what can this agent contribute, within what limits, and how does it interact?"

It is only an information model. There is no wire format, no schema, no media type and no protocol. Those are deferred to a future "IDL representation specification".

How It Fits In

Does this fit the agentproto charter? No.

The charter is narrow. It covers a dialog management protocol (dialog identifiers, lifecycle, context propagation), a reference architecture, and a use cases document. An agent description model is none of those.

Parts of it also run into the charter's out-of-scope list. The charter excludes "standardization of agent decision-making or planning semantics". IDL defines autonomy levels, escalation triggers, negotiation rounds and a six-phase intent lifecycle. The draft says repeatedly that these are descriptive metadata and not required behavior, which reads like an attempt to stay on the right side of that line.

The draft has no dependencies on IETF RFCs beyond BCP 14. It cites draft-rosenberg-aiproto-framework as motivation and draft-mcw-opsawg-icon-requirements as the governance framework it maps to. MCP and A2A appear as things IDL can sit beside or point at.

It also serves the author's own Agentic Intent Network architecture. Appendix C profiles IDL for AIN, and the agent class taxonomy (diagnostic, optimization, assurance) comes from network operations.

There is no reference to OAuth anywhere, even though the model overlaps heavily with it.

Summary

An IDL Description has six components.

  • Agent Identity: a URI, class, provider, roles, optional trust domain and delegation chain, plus credential-bindings that list the authenticated subjects the agent is expected to show up as.
  • Capability Declarations: capabilities named by opaque URIs, with optional input and output schemas and pointers to the MCP tools or OpenAPI operations behind them.
  • Autonomy Envelope: an autonomy level (advisory, supervised, bounded, autonomous), scope constraints such as a maximum transaction value, and an escalation policy.
  • Intent Participation: which intent classes the agent joins, in what roles, with which interaction patterns and negotiable parameters.
  • Context Model: what context the agent keeps, needs and will share.
  • Governance Interface: where a supervisor can call to read decision logs, pause the agent, or halt it.

So what does a receiver do with this? It authenticates the peer using whatever the communication protocol provides. It then checks that the authenticated principal matches one of the credential-bindings entries. If it matches, the receiver can use the description to select agents and plan the interaction.

Everything in the description is self-reported. The draft is upfront about this. The autonomy envelope "is not a grant of authority", and effective authority is the intersection of the envelope, the permissions of the authenticated principal, and external policy.

That means the actual authorization still lives somewhere else. In an OAuth deployment, it lives in the access token.

Interoperability

Significant interoperability concerns

There is nothing to implement yet. The JSON in Section 4 is explicitly informative. Two implementations cannot exchange an IDL Description based on this document.

The draft has normative requirements with nothing to hang them on. It has around 20 MUSTs. "Receivers MUST reject duplicate capability URIs" is a parsing rule for a format that doesn't exist. Section 8.4 requires that an agent "cannot repudiate a commitment made during negotiation" without defining a single negotiation message.

The value spaces are open. Scope constraint dimensions are enumerated carefully. Almost nothing else is.

  • credential-bindings.scheme uses oauth-client and authenticated-workload-identity in examples, with no definitions and no registry.
  • Escalation triggers and actions (confidence-below, pause-and-notify, negotiate-with-peer) appear only in examples.
  • participation-mode shows up as collaborative and lead-or-collaborate, with no defined list.
  • delegation-chain has no structure at all.
  • Operation classes like search and book are undefined strings.

This matters because the draft says an implementation that doesn't recognize a dimension, operator or identifier MUST treat the action as outside the envelope and escalate. Across vendors, with no shared vocabulary, everything escalates.

Capability and intent URIs are opaque with no shared vocabulary. Matching is exact string comparison. Two travel agents from different vendors only match if someone outside this spec coordinates the URIs.

The credential binding is weaker than it looks. The description lists its own expected subjects. Matching the session principal against that list stops one agent from presenting another agent's description. It does not make provider, trust-domain or the autonomy claims true. The draft says a protocol mapping must bind the description to an integrity mechanism, and leaves that to the mapping.

Overlap with OAuth

The autonomy envelope restates authorization concepts the OAuth working group already has.

A transaction-value maximum scoped to urn:travel:account:alice is what Rich Authorization Requests expresses in authorization_details. The difference is that RAR data is issued by the authorization server and enforced by the resource server. IDL's version is published by the agent about itself.

delegation-chain covers the same ground as the actor claims in Token Exchange, again without an issuer.

authorization-profile on governance bindings names a role like incident-commander with no way to say how a caller obtains it.

The risk is two descriptions of the same limit that drift apart. The draft's answer is "most restrictive wins", which is correct and also means the self-declared copy adds no enforcement value.

Mechanisms that could be removed

  • Agent class taxonomy. Six required classes that no receiver behavior depends on.
  • Top-level roles. Roles are already declared per intent class.
  • Lifecycle participation flags. Booleans like self-verify and progressive-reporting have no defined effect.
  • Context model. source, freshness and context-gap-behavior describe agent internals. The peer only needs to know what context is required and when.
  • Sections 5.4, 6.1 and Appendix B. Three comparison tables making the same positioning argument.
  • Section 7. Message flows with invented message names in a document that says it defines no messages.

The governance interface is the most concrete idea here. A pointer to "this is where you halt me, and this is the role you need" is small and testable. It could stand alone.

Writing Quality

The writing is clean and well organized. Sentences are grammatical and the document structure is easy to follow. It reads like a design proposal more than an IETF spec.

At 41 pages it is long for what it specifies. The complete JSON example runs seven pages, and several fragments appear twice.

There are some inconsistencies.

  • BCP 14 keywords are used throughout an Informational information model with no conformance target.
  • agent-id is described as identifying "this agent instance", but the example is https://agents.travel.example/planner/v2, which looks like a product version.
  • Table 16 lists governance as a sibling of its own child fields.
  • Section 3.4.3 gives priority and constraints on negotiable parameters. The complete example uses currency and range instead.
  • References are given without authors and with draft names in place of URLs.

Appendix B undersells A2A. It describes Agent Card skills as a "string list" and lists identity and security support as minimal. Agent Card skills are structured objects, and the card carries security scheme declarations. The comparison would be more convincing if it were accurate.

The central design choice lacks justification. Why would a peer make decisions based on an autonomy envelope the agent wrote about itself? The security considerations say not to rely on it. The draft never explains what a receiver should do with it instead.

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-session-requirements and draft-feng-dmsc-intent-routing-requirements.

The author has IETF experience, mainly in YANG and network operations. The agent work is recent and comes as a connected set: AIN architecture, NAIM, intent routing, and now IDL. This draft is best read as one piece of that larger proposal looking for a home.