Update review · November 22, 2026
"Agent DNS-OAuth Link -03: Client Migration Rules and a More Candid Implementation Report"
This is an update to the review of -02. The draft lets a receiver check that an agent's OAuth client hasn't been disabled, using a DNS TXT record and a no-permission access token from the agent's own AS.
What changed
Nothing on the wire changed again. The record, the link token, the Agent-Link-Name header and the receiver's steps are the same as -02.
The update is about two pages of clarifications:
- A new Section 8.2.2, Change of Issuer. The domain owner moves an agent to a new client or a new AS by editing the record. Receivers pick up the new values when their cached copy expires.
- Softer claims in the abstract and goals. Disabling the client now stops the agent "within a bounded time", at receivers "that follow this procedure".
- A note on backdated
iat. Some authorization servers setiata few minutes before the time of issue. A maximum age shorter than that offset refuses every token. - Removing the record now unlinks the agent. Once a receiver's lookup succeeds and shows no record, it stops requiring a token. -02 said removal only affected new receivers, which contradicted Section 8.1.
- A rationale for TXT over SVCB. The record describes the client an agent uses for outbound calls, so it applies even when the agent has no endpoint.
- One sentence on domain takeover. DNSSEC doesn't help if control of the zone is lost.
- Two additions to related work. Agent identities registered at an identity provider, and the DAWN working group on agent discovery.
- A longer implementation report.
No mechanisms were removed. There are no new authors.
Was the previous feedback addressed?
Two points were touched.
Revocation is no longer oversold. The abstract used to say disabling the client "stops the agent being accepted by every receiver". It now says within a bounded time. That matches what the mechanism does.
The caching question got half an answer. The last review asked in what scenario DNS caching extends the time a disabled client is accepted. Section 8.2.2 describes a real case where the DNS cache matters: the owner edits the record, and tokens for the old client are accepted until the cached record expires.
That is a different event from disabling the client. The formula in Section 8.2 still adds DNS and key caching periods to the delay after a disable. Section 8.2.1 still says increasing DNS caching lengthens the time a disabled client is accepted. Neither is true for a disable, where the record and the keys don't change. The fix is to split the two bounds: token age plus clock skew for a disable, record cache lifetime for a record change.
Everything else is still open:
- The token still has no defined place in the request. No header is named, and nothing says how it coexists with a real access token for the receiver's API.
- Channel mode is still not a protocol.
- "Refuse" is still undefined. No status code, error code or challenge.
- DNSSEC is still a SHOULD in sender-constrained mode.
- Once-linked-always-linked is still a SHOULD.
- SSRF is still not mentioned.
- The alternate client claims (
azp,appid,cid) are still allowed. - The
audfield is still in the record. - There is still no example HTTP request or token.
- RFC 8705 and DNSSEC are still informative references.
- The comparison to HTTP Message Signatures is unchanged.
- The test procedure in Section 8.2.1 is still in the body of the spec.
Does the new Change of Issuer section work?
It describes the mechanics correctly. It also exposes a conflict between two of the draft's own goals.
The section admits that "the record alone does not show whether a change was made by the domain owner". Its answer is Section 8.3: receivers that write policy on both the agent name and the client identifier won't carry permissions over to the new client.
So what happens in a legitimate migration? The operator moves agent01.a.example from one AS to another and updates the record. Every receiver that followed the SHOULD in 8.3 now sees an unknown client ID and denies the agent. Each partner has to update its policy by hand. The first goal of the draft is "no per-partner configuration".
Receivers that key policy on the name alone avoid that. For them, whoever can edit the record can point the name at their own client and inherit the agent's permissions. In sender-constrained mode there is no other check on the name.
The draft needs to pick one. Either the client ID is part of the agent's identity at receivers and migration is a per-partner event, or the name alone is the identity and DNS control is everything.
There is also a gap during the switch. The draft says tokens for the new client "may be refused" until caches expire, and suggests lowering the TTL first. A receiver could instead refresh the record when a token fails the client check. That would make migration close to instant and is cheap to specify.
Is the unlinking text right?
It is now consistent with Section 8.1. It is also worth reading next to the revocation story.
An operator disables a client. Tokens stop. If the record is later removed, receivers treat the agent as unlinked and go back to "normal policy" with no token required. The draft never says what normal policy should be for a name that used to be linked. A receiver that accepts unlinked callers from that name has just re-admitted a disabled agent.
Removal should come with guidance: a receiver that has seen a link for a name should treat unlinking as a policy change that needs review, not as a return to the default.
Implementation status
This section is the most useful change in -03. It is far more candid than before, and it shows how far the running code is from the spec.
The author's product, HexaEight Agent, now lists five differences:
- It uses a different record prefix at a different name.
- When several records exist, it uses the first. The draft says MUST refuse.
- It reads the client from
azp,appid,cid, thenclient_id. The draft says receivers MUST useclient_idwhen present. - Only the sender checks that the token grants nothing. The draft lists that as a receiver check, and any failure MUST result in refusal.
- Sender-constrained mode is not implemented.
Testing was done between two agents, both run by the implementer, with Microsoft Entra ID as the AS. Entra access tokens identify the client with appid or azp, not client_id. The only tested path is the "compatibility allowance", not the RFC 9068 path the draft calls normal.
The reported tests include refusal "once the client was disabled and the sender could obtain no new token". That shows token issuance stops. It does not report the measurement Section 8.2.1 asks every receiver to run, which is replaying tokens issued before the disable and timing the last acceptance.
The HTTP-facing mode, with the Agent-Link-Name header and DPoP, still has no running code.
Where this leaves the draft
Four revisions in a month. The text keeps getting more honest about limits, which is good. The protocol still can't be implemented by a second party.
The questions for the working group from the earlier reviews all still apply:
- Is presenting an access token to a party that is not its audience a pattern OAuth should standardize?
- Does this beat publishing a key in DNS and using HTTP Message Signatures, given both rest entirely on DNS?
- Where does the token go in the HTTP request, and how does it coexist with a DPoP-bound access token for the receiver's own API?
- Should DNSSEC be a MUST for sender-constrained mode?
Two new questions for the author:
- Is the client ID part of the agent's identity at receivers? If yes, how does migration avoid per-partner reconfiguration?
- Since the only tested AS doesn't emit
client_id, should the draft define the client claim per issuer instead of leaving it to each receiver?