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

Review · August 21, 2026

Cross-Organizational Delegation for Workload and Agent Identity: Problem Statement and Requirements

draft-reece-wimse-cross-org-delegation -01View on Datatracker →

Problem Statement

This draft addresses the authorization challenge that arises when autonomous AI agents delegate sub-tasks across organizational boundaries — for example, an orchestrator agent at Company A invoking a tool at Company B on behalf of a human at Company A. Existing workload identity and token mechanisms (including OAuth and RFC 8693 token exchange) were designed for shallow, intra-organization delegation to a known authorization server. This draft argues they cannot handle: recursive authority attenuation across multiple hops, cross-org verification without pre-provisioned federation, and offline verification without per-request callbacks.

This is a problem statement and requirements document only — it does not propose a solution.

How the Spec Fits In

This draft sits at the intersection of the WIMSE and OAuth working groups, extending concerns that WIMSE-ARCH has identified but not fully resolved. Rather than defining new mechanisms, it articulates requirements that solution-oriented drafts (in WIMSE or elsewhere) would need to satisfy.

It references only two normative sources (RFC 2119 / RFC 8174 for key word usage) and one informative source (draft-ietf-wimse-arch-08), making it self-contained but lightly grounded in prior IETF work. It does not extend any existing RFC — it evaluates gaps in several (OAuth session model, RFC 8693 token exchange, workload identity mechanisms) at a high level.

This document does not establish new relationships between OAuth roles or define new protocol interactions. It is framing work.

High-Level Summary

The draft enumerates seven problem facets that occur when an AI agent chain (human → orchestrator agent → sub-agent → tool) crosses organizational lines:

  • Recursive attenuation: Authority should narrow at each delegation hop, but today's bearer tokens don't verifiably enforce this.
  • No pre-existing federation: In open, many-to-many agent deployments, requiring bilateral federation between every pair of organizations is impractical.
  • Offline verification: High-frequency tool calls can't afford per-call runtime callbacks to the originating org's authorization server.
  • Principal binding: The human's identity and entitlements must travel through the chain in a way no intermediary agent can silently alter (the "confused deputy" problem).
  • Revocation: Revocation information must reach cross-org relying parties without requiring a live callback, at bounded staleness.
  • Composable audit: Each org sees only its slice of the delegation chain; records need to be tamper-evident and composable for incident investigation.
  • Execution-time human authorization: Some actions (payments, deletions, safety boundary changes) require a binding human decision at the moment of execution — not just standing authority granted earlier.

The ten resulting requirements (R1–R10) are stated mechanism-independently: no specific token format, signature scheme, or policy language is mandated. R10, added in -01, is the most novel — it requires that any designation for human approval survive delegation (can't be stripped), that approval evidence bind to specific action arguments, that evidence be consumed at most once, and that unrecognized designations cause the request to be denied.

Interoperability and Necessity

Assessment of scope: This document is a problem statement, so the normal "is enough specified for interoperability?" bar doesn't directly apply. The more relevant question is whether its requirements are precise enough to guide future solution specs toward interoperable outcomes. On that measure, the draft is mixed.

Requirements R1–R7 and R9 are reasonably well-scoped for a requirements doc and would usefully constrain solution proposals. They are consistent with how the space is discussed in WIMSE-ARCH and related work.

R10 is significantly under-specified for a MUST-level requirement. Terms like "execution-time human authorization," "designated," and "at-most-once reliance" are introduced but not defined with enough precision to evaluate competing solution approaches. The security considerations note that at-most-once in replicated deployments is an unsolved distributed-systems problem — flagging this as a MUST requirement without even sketching the design space risks making R10 unsatisfiable in practice. It would be stronger as a SHOULD with an explicit acknowledgment of the open problem.

Mechanisms that could be removed or deferred: The composable audit requirement (R8) is already a SHOULD, which is appropriate given how poorly specified it is. It could reasonably be dropped from this version entirely, or split into a separate problem statement, since it introduces significant complexity (tamper evidence, cross-org record composition) that is largely orthogonal to the authorization problem.

Writing Quality

The draft is well-structured and uses correct IETF key word conventions (RFC 2119/8174 cited). Sentences are grammatically sound and generally precise. The problem facets in Section 3 are clearly separated and logically ordered.

The draft is appropriately concise for a problem statement. The seven-facet structure is a useful organizing principle and the requirements table format aids readability.

One notable quality concern: The Non-Goals section explicitly disclaims selecting "credential formats, signature schemes, or policy languages," but R10 introduces behaviors (at-most-once evidence consumption, binding to specific action arguments) that implicitly constrain what formats and schemes are viable. This tension between the non-goals and the requirements is not acknowledged.

The security considerations section is more substantive than typical for a requirements draft, which is a positive signal that the author has thought carefully about the problem. The R3/R7 tension (offline verification vs. timely revocation) and R5 binding concerns are well-articulated.

Background of the Authors

Morgan Reece ([email protected]), TowerGuardian Consulting, Austin TX

  • RFCs published: 0
  • Active Internet-Drafts: 1 (this document)
  • Assessment: New author in the IETF with no prior publication history. The single active draft is this one, in -01 revision.

The limited IETF track record means there is no prior record to evaluate writing style or follow-through. TowerGuardian Consulting does not appear in other IETF-associated organizations. The quality of this draft is better than many first-time submissions — the problem framing is coherent and the requirements are largely well-motivated — but the novelty of R10 and the lack of existing-work engagement (only one informative reference beyond boilerplate) suggest the author may benefit from more engagement with adjacent work in OAuth (transaction authorization, RAR) and WIMSE before a working group adoption discussion.