Update review · September 2, 2026
Agent Delegation Profile Update
Two days after -00, this draft went from 28 pages to 48. The author answered one of the four questions from the previous review, added the asor comparison the WG was going to have to write itself, and built two new mechanisms on top of the three foundational gaps that are still open.
What got fixed
The closed axis set is now a registry. Section 9.1 defines the "HAMR Floor Axis Registry," Specification Required, with four comparators: min, max, rank, one_of. Every registered axis name binds to exactly one comparator. An unregistered axis name is a hard rejection, and so is a registration naming a comparator outside the set of four. Eight initial entries, and the -00 axes are folded in without changing anything on the wire: the three enum axes register as one_of with singleton sets, which is what their equality-only rule already was.
This directly answers the -00 question about what a non-telco deployment puts in a floor. The answer is now "register an axis," and three of the eight initial entries are things anyone can use.
The asor comparison exists now. Section 17 has a real paragraph on draft-asor-wimse-agent-delegation-chain. asor's registry names comparator types (max, one_of, not_one_of, prefix, rank) and leaves the constrained quantity to a profile. This one names axes, each bound to one comparator. The draft calls the comparator vocabulary the bridge between the two, notes asor has no min entry because every asor constraint narrows a ceiling while these duration axes tighten upward, and says a mirrored min "has been stated as intended for a future revision of that draft."
That last sentence means the two authors have talked. The coordination the -00 review asked the chairs to force is happening without them. Worth noting, and it doesn't settle anything: the two drafts still have incompatible architectures, and asor's registry includes a prefix comparator that this draft's Section 7 explicitly forbids for scope.
Section 4 places the draft among neighbors. Three layers: who may act on whose behalf, what exactly may be done per invocation, what happened once. This document is layer (a) plus floors. RAR binds detail at grant time, draft-das-agentic-tool-binding binds a tool call's arguments per invocation, draft-schrock-ep-authorization-receipts records what happened. RFC 9396 is the first OAuth specification this draft has ever cited.
The section also names the parent-still-valid problem: attenuating a credential into a narrower child doesn't invalidate the parent, so the parent's holder keeps the broader authority. It cites asor 9.1 for mitigations and says plainly that it has none.
Still no RFC 8693, and still no mention of the act claim. That gap was the main criticism of -00's positioning and it survives untouched.
The new mechanisms
Action class floors (Section 10). actionClass is r < w < x: read-only, idempotent write, non-idempotent consequential action. It's a ceiling on the link, and a child can only lower it. classSource is method < declared and says where the verifier gets a request's class from. method means classify from the HTTP method alone. declared means additionally consult a menu the Resource Owner signed. method is the tighter value because it never trusts a class asserted by anyone but the verifier, so a child link can move from declared down to method and never the other way.
The declared menu is a JWS whose payload carries an origin and a map from "METHOD path-template" to a class. The matching rule is spelled out segment by segment, longest-literal-match wins, and a tie is a miss rather than a coin flip on JSON key order. The menu's iss must match the request target's origin octet-for-octet, the key comes from a configured trust source and never from the menu, and any menu failure falls back to the method default rather than combining the two.
This is the most carefully specified part of the document. The origin binding, the no-combining rule, the tie-is-a-miss rule, and the V9 negative control are all the right instincts.
Write budget (Section 11). writeBudget is a non-negative integer with a max comparator, cumulative across the chain, counting every admitted w or x operation. The draft separates the limit, which rides on the wire, from the count, which is verifier-held state keyed by a chain identifier. Section 12.2 is cited as the reason the count can never travel: anything on the wire can be re-presented, and a count that can be re-presented is not a count. The chain identifier is the SHA-256 digest of L(0)'s signature value, derived by the verifier, never taken from the request. Section 11.4 admits that N verifiers with no shared state means N budgets.
These two axes retroactively justify rule 2. The -00 review asked why a sub-agent would raise a required SIM age. It wouldn't. A parent agent handing a sub-agent read-only and a smaller budget is the obvious case, and now the draft has axes where child-tightening is the natural thing to do.
The three gaps that are still open
Nothing in -01 touches any of these.
No link commits to its parent. Section 12's required attestation properties still don't include a parent reference. Section 8 still establishes parent-child purely by position in the header list. Splice L(0) from a chain you hold with L(1) from a chain someone else presented, order them so the three rules pass, and both signatures verify because both links are genuine.
This now has a second payoff. The write budget is keyed on L(0)'s signature. A presenter with an exhausted chain can splice in a different valid L(0) that satisfies attenuation against their existing leaf and get a full budget back.
No link commits to the presenter. Still a bearer credential. Capture the header bytes, sign your own request with your own key, and the chain verifies.
The link encoding is still undefined. Section 5 still says the credential format decides and this document doesn't mandate one. That was an interop problem in -00. In -01 it's inside a security computation: Section 11.3 digests L(0)'s signature value "in whatever encoding that signature value is carried on the wire by the credential format in use." Two verifiers that pick different encodings compute different chain identifiers for the same chain.
The nonce requirement is still impossible. Step 7 still requires every floor attestation in the chain to echo the nonce the verifier issued for this exchange. The root attestation is minted on the person's device before the rental service exists. The worked example still shows no nonce.
New problems worth raising
A fresh root mints a fresh budget. Section 11.3 closes re-delegation carefully: a new child link doesn't change L(0), so it doesn't change the identifier. It closes caller-supplied identifiers, and V11 is a negative control for exactly that. It doesn't close re-rooting. Nothing stops the Delegator from signing a second L(0) for the same authority. New signature, new digest, new identifier, full budget. The cap binds to a signature, not to a person.
The draft never says which link's writeBudget is the limit. The count is one number per chain, keyed off L(0). The limit "rides on the wire, on every link that carries the axis," and every link can carry a different value. Presumably the leaf governs, since child values are always lower. That needs saying. It matters because two sub-agents descending from the same L(0) share one chain identifier and therefore one count, while carrying different leaf limits. Whoever spends first depletes the other, and which limit the shared count is checked against depends on which chain arrived.
Budget is spent before the relying service gets a vote. Step 9 classifies and decrements. Step 10 applies the Relying Service's own authorization policy. Nothing restores the count when step 10 refuses. A request the service was always going to deny still costs the chain a write. Either move the decrement after policy or require rollback.
The method table has no default of last resort. GET, HEAD, OPTIONS map to r. PUT, DELETE map to w. POST, PATCH map to x. TRACE is absent even though the draft cites RFC 9110's safe-method list, and CONNECT, PROPFIND, and every extension method have no defined class at all. The draft is emphatic that an unknown axis name is a hard rejection. An unknown method gets nothing. It should default to x.
The menu pins Ed25519 with no agility. "This document defines no algorithm negotiation or agility for the menu, and a menu using any other algorithm does not verify." A standards-track document with one hardcoded signature algorithm and no negotiation will not survive review.
The declared menu can make writes free. A Resource Owner declaring a POST as r means that operation never spends against writeBudget, because Section 11.2 exempts r-class requests. The two mechanisms are coupled and the draft doesn't say so.
The CAMARA numbers argue against the default. Section 10.7 reports 57 of 138 POST operations in the CAMARA catalogue are actually reads and get classed x under method. That's 41% wrong for the author's own target deployment, and the fix requires every Resource Owner to publish and sign a menu. The draft states this rather than hiding it, which is characteristic, and the WG should still ask whether a default that wrong is the right default.
The registry name. "HAMR Floor Axis Registry" names an IANA registry after the author's draft prefix. Registries get names like "OAuth Parameters." This one signals the author is planning for the document to stay an individual submission.
What still hasn't been removed
partialPolicy is still there, and it's now an IANA registry entry, which makes it harder to remove.
Section 12.1's requirement that a signed refusal be verified "along a verification path independent of the path used to verify an accepted attestation" is still an instruction about code structure that nobody on the wire can check.
Appendix A didn't get cut. It grew. Direction 4 adds three GitHub repositories that are explicitly not implementations of this profile.
RFC 8941 is still the normative Structured Fields reference. RFC 9651 obsoleted it in 2024.
"Published floor" is still undefined. It now carries more weight than it did in -00: the link-to-published-floor branch of the omitted-axis rule and writeBudget's effective-zero default both depend on it, and Section 2's terminology list doesn't include it. There's no statement of how a published floor is expressed, fetched, or authenticated. Section 9.1 also introduces effective as a name for a verifier's evaluated outcome and then says conveying it is out of scope, which leaves a defined term with nothing to do.
Writing quality
The new normative sections are good. Section 10.4's path-template matching rule is deterministic and closes the tie case. Section 11.2's separation of limit from count is a clean piece of thinking, and the reason given for why the count can't travel is correct.
Both habits from -00 got worse.
The hedging is now arguing with reviewers who did show up. Section 4 exists "so a reader does not mistake the boundary this document draws for an oversight." Section 11.4's limitation "is stated here rather than hidden." Section 19's is "not a defect to be concealed." Section 10.2 ends with a paragraph explaining that the author's own source code lists two values in the opposite order from the spec text and that a reader "should not read that ordering as a disagreement about which value is tighter." That belongs in a code comment.
The restatement is worse because there's more to restate. Appendix D describes six changes in about two and a half pages, and one bullet runs twenty lines to say a section was added. The -00 review said 28 pages could be twelve. It's 48 now.
About the author
Amr Hassan, Independent, San Francisco. Still sole author, no changes to the author list. Zero published RFCs, one active Internet-Draft, no expired drafts.
Two revisions in three days with a substantive registry addition and two new specified mechanisms is a fast response to review for a first-time author.
Where the working group should push
The registry answers the extensibility question and does it in a way that's compatible with asor's comparator vocabulary. That's real progress and it's the part worth keeping.
The three questions from the -00 review are unchanged and they now have more built on top of them:
- How does a link commit to its parent, and how does it commit to the Delegate it grants to? Chain splicing now buys a fresh write budget, not just unearned scope.
- What encoding does a link use? The chain identifier is a SHA-256 digest over an undefined byte string.
- How does the root attestation carry the final verifier's nonce, when the root exists before the verifier does?
Two new ones:
- What stops a Delegator from minting a fresh L(0) and a fresh budget? Re-delegation is closed and re-rooting isn't.
- Which link's
writeBudgetis the limit the shared per-chain count is checked against?
For the chairs, the call from -00 hasn't changed, and it's now easier to make. The two authors are talking about comparator names on their own. What they can't sort out between themselves is that both drafts are requesting IANA registries for overlapping purposes in the same problem space, and one of them is named after a draft prefix.