Dear OAuth Working Group,

I would appreciate technical clarification concerning the boundary between
OAuth authorization and the final effectuation of an autonomous agent’s
action.

RFC 9396 can express fine-grained authorization details. RFC 9449 provides
sender-constrained access tokens. The Transaction Tokens work propagates
identity, authorization, and request context through a call chain. These
mechanisms provide important inputs to secure agentic workflows.

My question concerns the remaining transition from authorized context to
the exact externally effective consequence.

Consider two examples:

● An AI agent is authorized to pay a specified beneficiary, but a
downstream transformation, compromised worker, retry process, or stale
queue entry changes the beneficiary, amount, or execution conditions after
an earlier authorization check.

● An AI agent is authorized to SEND a particular file to a specified
recipient, but the attachment, recipient, destination service, sharing
permissions, or link target changes before release.

At the final resource server or irreversible boundary, does any current
OAuth specification establish the following complete invariant?

1. The resource server reconstructs the actual semantic act from the
parameters that the ledger, egress mechanism, storage system, or actuator
will use.

2. All consequence-relevant fields are compared with the granted
authorization or transaction context.

3. Current protected revocation, generation, expiry, quota,
resource-version, and replay state are revalidated.

4. Any single-use or replay-sensitive authority is consumed atomically with
the exact resulting state change.

5. Crash recovery can distinguish committed, denied, uncommitted, and
indeterminate outcomes without duplicating an irreversible consequence.

6. Every path capable of producing the protected consequence is
equivalently mediated.

If a resource server performs all these functions at the last preventable
boundary, I regard it as functionally equivalent to the execution-finality
architecture described in my draft, regardless of whether the
implementation uses that terminology.

The narrower question is therefore not whether OAuth can be deployed at the
final boundary. It clearly can. The question is whether token validation at
that location automatically provides these semantics, or whether they
remain application-specific properties outside the OAuth framework.

I would particularly welcome clarification on the following:

1. Is there an existing OAuth specification or profile that already defines
this complete check-consume-effect invariant?

2. Are Transaction Tokens intended only to preserve authorization and
request context across a trusted-domain call chain, or are they also
intended to guarantee reconstruction of the final semantic act and atomic
coupling to the resulting external consequence?

3. If the latter properties are not currently specified, would an
execution-bound profile or informational requirements document be useful to
distinguish what OAuth carries from what the final resource server must
enforce?

The architectural analysis and a detailed FAQ addressing the “OAuth at the
last irreversible boundary” question are available here:

https://datatracker.ietf.org/doc/draft-das-execution-finality-deployment/

Title: “Execution-Finality Architecture for AI and Autonomous Critical
Systems”

I am not suggesting that OAuth, RAR, DPoP, or Transaction Tokens are
defective. I am trying to determine whether the end-to-end invariant is
already specified, remains entirely application-specific, or represents a
profile-level interoperability gap for autonomous-agent deployments.

References to existing specifications or implementations providing
functional equivalence would be especially helpful.

Regards,

Sangam Kumar Das
Independent Researcher
_______________________________________________
OAuth mailing list -- [email protected]
To unsubscribe send an email to [email protected]

Reply via email to