Review · November 16, 2026
Bounded Work Authorization, a Requirements Document for the Mission and Schemas Drafts
The Problem
OAuth can revoke a token. It has no shared way to say "the task the user approved is over."
An agent doing one job collects many tokens through code exchange, refresh, and token exchange. Nothing ties them to the one approval. Section 1.1 says an operation can be denied when the agent's identity and token are both valid, because the work has ended.
This is not a protocol. The abstract says it "specifies no wire formats and requests no IANA actions" and calls itself "a direction-setting document." It uses no BCP 14 keywords (Section 1.3).
How It Fits In
It is an informational framework. The only normative reference is RFC 6749. It defines no parameters, claims, endpoints, or metadata.
It still describes a lot of new behavior for whatever spec implements it. Section 4.4 has the resource server check each operation against the work's lifecycle and bounds. Section 4.3 lets the subject, the approver, or an administrator end the work at the AS without holding any token. Token revocation in RFC 7009 is client-facing, so that is a new relationship. The approver can also be a different person from the subject (Section 2).
So how does it relate to the authors' other drafts? It sits above them and replaces none of them.
- Mission-Bound Authorization is described in Section 4.5 as "one realization." The proposal maps to the Mission Intent, the governing authorization to the Mission plus its Authority Set, and approval evidence to Mission's integrity commitments.
- Authorization Schemas is the other realization. An approved schema instance with a work identifier is the governing authorization.
- Authorization Evidence is the third object. Section 4.5 notes its signature covers the displayed content and not the enforceable bounds.
Two authors with competing designs have written down the part they agree on. That is useful for the working group.
Summary
There are three objects (Section 4.1).
- Bounded work proposal. The agent describes the work, the bounds, and when it ends. It confers no authority.
- Governing authorization. The record the AS creates on approval. It has its own identifier, an immutable approval record, and a state of active or ended.
- Approval evidence. An AS-signed record of the decision, kept after the work is over.
Three invariants connect them (Section 4.2). Approval sets a ceiling that never rises. Every derived credential stays inside the ceiling and stays associated with the work, including tokens a sub-agent gets through Token Exchange. Evidence is never current authority.
Ending is terminal and has five reasons: completed, abandoned, withdrawn, terminated, expired. After it, the AS refuses every issuance path. Asking for more authority creates a new governing authorization and never edits the old one.
Section 5.1 reduces this to nine requirements, R1 through R9. Section 5.3 gives a test scenario any candidate mechanism can be run against. This is the best part of the document.
Interoperability
Nothing here can be implemented, by design. Are the requirements tight enough to choose between realizations?
Significant interoperability concerns
The acceptance window can be the token lifetime. In the walkthrough, Alice withdraws at 22:00 and the agent is denied at 23:00. That only happens because both resource servers introspect every token. Section 3.6 says a deployment relying on token lifetime alone "still conforms." R6 is met by any AS that states a number.
The "no changes" claim conflicts with R9. Section 5.1 says a realization "works with existing clients and authorization requests" and a resource server "applies an ending without change." R9 has the resource server refuse any token with no association to a governing authorization and publish which operations need one. That is new resource server behavior. The draft also doesn't say how an existing authorization request carries a proposal's termination conditions.
Narrowing has no comparison rule. R3 needs a subset check per authority type. Section 4.2 admits authorization details "have no standardized comparison" and tells the AS to refuse when it can't decide. The $600 versus $500 test in Section 5.3 only passes for types someone has defined.
The second authorization server has no way to learn anything. Section 4.3 says a second AS refuses issuance "once it learns of the ending." No mechanism is named. A separate identity provider "learns of the ending as a second Authorization Server does."
Error reporting departs from two RFCs. Section 4.4 wants the agent to tell "work ended" from "out of bounds" from "token failed." The draft says this departs from RFC 7662 Section 2.2 and the invalid_token retry in RFC 6750.
Withdrawal still has no interface. R5 names the requirement. Our Mission review flagged the same gap, and this draft does not close it.
Mechanisms that could be removed
- Approval evidence. Section 4.1.3 says no party needs it to decide whether an operation is permitted, and a realization can skip it. None of R1 through R9 requires it.
- Five ending reasons. Section 4.3 leaves who may report completion to the deployment. The enforcement outcome is the same for all five.
- Standing approvals and work across two authorization servers. Both can wait. A first version could cover one consent screen and one AS.
Writing Quality
The writing is clear and well organized. The walkthrough in Section 3 is easy to follow. Quotes from other specs come with section numbers.
It is repetitive. The evidence-is-not-authority point appears in Sections 3.2, 3.5, 4.1.3, 4.2, and 7. The two-part resource server question is stated in full in 3.3 and again in 4.4. At 36 pages it is long for a problem statement.
Several references are stale. It cites Mission -00 when -02 is current, Authorization Evidence -01 when -02 is current, and ID-JAG -04 when -06 is current.
Authors
From Datatracker, fetched today:
- Dapeng Liu (Alibaba). 3 RFCs from 2014 and 2015, all in mobility management. 14 active drafts and 77 expired. An experienced author writing a lot right now.
- Karl McGuinness (Independent). 0 RFCs, 25 active drafts, none expired. Co-author of ID-JAG, an adopted WG document. Writing at a very high rate.
Questions for the Working Group
This is on the Friday agenda. Section 1.3 asks three questions, and the first is the one to answer in the room.
- Should OAuth standardize the link between approved work and the tokens issued for it? If yes, is that an OAuth WG item or a job for a profile of Grant Management?
- Is R1 through R9 the right list? Which requirements does the WG reject? R9 and the multi-AS rule are the expensive ones.
- Should R6 set a maximum acceptance window, or is "state your token lifetime" enough?
- Is this document meant for adoption, or is it input to choosing between Mission and Authorization Schemas? Will the authors commit to one?