A revised Internet-Draft published on 16 November proposes two email header fields that declare a message came from a computer agent and name the humans responsible for it. The revision tightens the authentication rules: the headers must be DKIM-signed, and the responsible humans’ domains must align under DMARC.
What happened
W. Chuang (Google) and E. Gustafsson published revision -01 of draft-chuang-agent-attribution-header-fields, “Agent Attribution Header Fields”. It is an Experimental Internet-Draft in the Independent Stream; revision -00 was published on 19 June 2026. The document has no working-group adoption and, as published, no IANA actions.
The draft defines two header fields. Agent-of carries an address-list of the humans responsible for the agent that sent the message. It is meant to be both human and machine readable, so a recipient who sees unwanted mail can write to a person rather than to a bounce address. The address does not have to be a primary mailbox; an alias or a mailing list is allowed, provided it can be reached over SMTP.
Agent-attribution carries base64-encoded JSON and is machine-readable only. It is the extension point for complex responsibility and delegation relationships — an agent acting for a team, or an agent that created another agent. The JSON keys are not defined yet; the draft says it expects an IANA registry for the values and future updates to describe the relationships. When an agent creates another agent, the responsible human information must be propagated if the new agent is to use the header, and it must include a means of reaching that human over SMTP.
If both headers are present, a receiver should take the union of the identities they name.
The authentication rules are the substance of this revision. Senders must sign the headers with DKIM (RFC 6376), not merely should. Receivers that use the headers must DKIM-authenticate them: the signature must pass, and the agent header fields must be covered by the header hash. Where DKIM2 is available, its results should be used instead, with the same coverage requirement. To stop someone from naming a responsible human who never agreed, receivers must check DMARC relaxed alignment (RFC 9989) of the human identifiers with a passing DKIM or DKIM2 signature domain.
The draft is careful about what the headers mean. Their presence does not imply trustworthiness: they are a claim by the sender that the message came from an agent and a claim about who is responsible, plus an assertion that the named humans can be reached. The Agent-of field says nothing about the authentication requirements of the agent or of the From header, because — in the draft’s words — it is “solely about agent attribution as opposed to agent identity“.
Why it matters for KYA
Email is the first place most agents act in public. Outreach, support replies, notifications and agent-to-agent correspondence all arrive as mail, and the receiving side currently has no standard way to tell an agent-sent message from a human-sent one, or to find out who to contact when the agent misbehaves. That gap is why abuse teams treat agent mail as either spam or an unmanageable exception.
This draft is the minimum viable version of KYA for that channel: an attributable principal. It does not verify the agent, bind a key to it, or say what it is allowed to do. It gives the recipient a claim — signed by a domain, aligned under DMARC — that a named human is accountable and reachable. That is less than an identity system, and more than most mailboxes have today.
The DKIM and DMARC requirements are what make it worth reading. Anyone can type an Agent-of header naming a bank’s security team. Only the domain that passes DKIM can make that claim stick, and only DMARC alignment stops a lookalike domain from naming someone else’s employees. The revision’s move from SHOULD to MUST on signing is the difference between a header that filters can trust and one they must ignore.
The limits are equally useful for buyers and for mail administrators. Presence of the header is not proof of good behaviour; it is a contactable claim. The machine-readable header has no schema yet, so interoperability beyond the human-readable address is not there. And because this is an Experimental draft in the Independent Stream, no mailbox provider is obliged to do anything with it.
What to watch
- Whether a large mailbox provider implements the headers, or publishes guidance on filtering them. Google authorship is a signal of intent, not a product commitment.
- Whether the JSON schema for
Agent-attributionis defined. Without it, the machine-readable half is a container with no agreed contents, and delegation chains cannot be checked. - Whether the draft is adopted by an email standards group. Attribution headers that mailbox providers do not implement remain a vendor convention.
- Whether agent platforms add the headers before receivers use them. As with every sender-side signal, adoption is a coordination problem: the value only appears when both ends ship.
Sources
- W. Chuang and E. Gustafsson, “Agent Attribution Header Fields”, draft-chuang-agent-attribution-header-fields-01, 16 November 2026 — datatracker.ietf.org
- W. Chuang and E. Gustafsson, “Agent Attribution Header Fields”, draft-chuang-agent-attribution-header-fields-00, 19 June 2026 — datatracker.ietf.org
- D. Crocker, T. Hansen and M. Kucherawy, “DomainKeys Identified Mail (DKIM) Signatures”, RFC 6376 — rfc-editor.org
- T. Herr and J. Levine, “Domain-Based Message Authentication, Reporting, and Conformance (DMARC)”, RFC 9989 — rfc-editor.org
Know Your Agent (KYA) explains agent identity, verification and accountability. This is an explainer, not legal or compliance advice — see our Sources & methodology. New to KYA? Start with What is Know Your Agent? and the glossary.