Flagged 6 November 2026 for review before the meeting (14 to 20 November 2026) · All session times PST (UTC-8) · Summaries written from the draft text, read through the ietf draft CLI
The draft takes the loose term "machine readable" and splits it into two core criteria that belong to the expression itself, plus three orthogonal dimensions that belong to the transaction. Parseable means a grammar yields the same structure for every conforming agent; interpretable means each term is pinned to an identified, shared vocabulary such as AIPREF's train-ai, unlike a guessable field like may_train. Discoverable, actionable (the expression settles this agent's pending decision without a human) and verifiable (origin and integrity can be established) can each hold or fail independently. Assessment gives a profile rather than a score, and the draft states there is no conformance threshold. It spends most of its length on legal terms of service: terms wrapped in a JSON string are parseable only as an opaque string and are actionable for no agent at crawl scale. It explicitly excludes interpretation by a language model as a way of satisfying any property, because a non-deterministic process determines nothing. A survey places robots.txt, AIPREF, ODRL, DPV, CC REL, TDMRep, Schema.org, Dublin Core, C2PA and FAIR on the five properties, noting that AIPREF reaches interpretability but offers no verifiability. It defines no vocabulary, syntax or protocol, proposes no change to existing mechanisms, and takes no position on whether anything meets the EU DSM Article 4(3) "machine-readable means" test.
The five properties are a ready checklist for agent directory and Agent Resource Discovery records in DAWN: a card can be discoverable and parseable while still not interpretable or verifiable. It also names the exact gap between AIPREF (interpretable) and Web Bot Auth signatures (verifiable) that his discovery work has to bridge.
IETF 127: AIPREF Mon 16 Nov 11:30 PST and Thu 19 Nov 11:00 PST; GEN area Wed 18 Nov 09:00 PST. Mon 11:30 overlaps NMRG; Wed 09:00 overlaps OAuth session 1.
Four drafts read together: the WG protocol, the earlier key directory spec, the Signature Agent Card and registry, and the intermediary principles.
A profile of RFC 9421 in which every bot request is signed with a key owned by its provider and carries a Signature-Agent header naming the HTTPS URL where those keys are published as a JWK Set. The identifier is the URL, not the key, so keys rotate without losing continuity; when no URL resolves, the verifier falls back to the key thumbprint, which cannot rotate. The well-known URI http-message-signatures-directory lets the URL name a domain, and optional possession proofs on the directory response bind the keys to it. Origins can ask for a signature with Accept-Signature, and the draft carries deployment guidance on caching, bounded fetches, SSRF and rollout. It is explicit that a valid signature shows only that a holder of a key the URL publishes signed the message: nothing about who operates the agent, whether it is benign, or whether the request is authorised. Human authentication, anonymous authentication, authorisation, delegation and sharing reputation between origins are out of scope. Revision -01 made every Signature-Agent value a direct JWK Set URI and removed the directory, jwks_uri and cimd types.
The narrower, earlier key directory specification: a JWKS restricted to registered HTTP signature algorithms, served at the well-known URI with media type application/http-message-signatures-directory+json. It defines Signature-Agent as a Structured Fields dictionary whose members are typed as directory, jwks_uri or cimd, with data: URIs allowed for inline, ephemeral keys. Directory responses should carry one signature per key covering @authority and content-digest, tagged http-message-signatures-directory, so a client can bind the key set to the origin. Delegation through x5c chains, AIA or x5u appears in examples but is flagged experimental. Most of this text now lives in the protocol draft, which has since dropped the typed members.
Defines the Signature Agent Card, a JSON document that reuses the OAuth Client ID Metadata Document namespace (client_id, client_name, contacts, jwks_uri or jwks) and adds one web_bot_auth object, so any existing CIMD is already a valid card. The extension carries expected User-Agent, robots.txt product token and a compliance flag, trigger (fetcher or crawler), purpose, targeted content, rate control and rate expectation, known URLs and an IP list URI, under a new IANA registry. A registry is a plain text list of card URLs, key directory URLs or inline data URLs, optionally access-controlled and polled with conditional requests. Keys referenced by a card must carry possession proofs, which is stricter than the protocol draft. Several members are still TODO (purpose vocabulary, rate formats, signing inline jwks), and purpose points to the AIPREF vocabulary as one option; the card is still referenced through type=cimd, which protocol -01 no longer defines.
Five principles for CDNs, security services and edge platforms: classify by behaviour rather than client identity, offer graduated responses instead of block or allow, base decisions on open protocols and standard status codes (403, 429, 503 with Retry-After), prefer active signals (HTTP Message Signatures, Privacy Pass, published IP ranges, forward-confirmed reverse DNS) over passive fingerprints, and document where services diverge from published standards. It states that a client representing a user implies neither unconditional access nor an override of site policy, and that the absence of an active signal does not establish abuse. Authentication is said not to establish authorisation, and identity signals are flagged as a tracking risk. There is no protocol and no IANA action, and acknowledgments are still TODO.
flowchart LR U["User"] --> A["Signing agent"] A -- "publishes keys" --> D["Key directory: JWKS at /.well-known/http-message-signatures-directory"] A -- "publishes metadata" --> C["Signature Agent Card: CIMD fields + web_bot_auth"] C -- "jwks_uri" --> D A -- "signed request + Signature-Agent URL" --> O["Origin verifier"] O -- "resolves URL, fetches JWKS, checks proofs" --> D R["Registry: list of card and directory URLs"] -.-> O R -.-> C O -- "graduated response: allow, 429, 503, challenge, 403" --> A
web_bot_auth) whose jwks_uri points at the directory.Signature-Agent with the key URL.The Signature Agent Card is an agent card at an HTTPS URL carrying metadata and keys, which overlaps directly with the agent directory records in DAWN; picking CIMD as the base namespace is a precedent DAWN should either reuse or explicitly diverge from. The "URL is the identifier, nothing more" trust model is also the floor that agent identity in WIMSE and OAuth will be measured against.
type=cimd, how does an origin get from a Signature-Agent URL to a Signature Agent Card, and will the registry and directory drafts be aligned or retired?IETF 127: WebBotAuth Tue 17 Nov 16:30 PST; bcp-civil also at the WIT area "AI and WIT" item, Wed 18 Nov 11:30 PST. WIT area overlaps BPM.
Selo targets the per-pair connector problem between organizations' systems, including MCP tool servers, which scale with the number of pairs and carry long-lived shared credentials. Agents owned by verified legal entities verify each other, disclose catalogs on a need-to-know basis, negotiate a typed access agreement that both sides sign, and then execute it through a gateway in front of each side's existing API. Each entity is anchored to a legal identifier (such as an LEI) by a name-constrained entity CA certificate, and issuance, key changes and revocations go to a CT-style or SCITT log that never records relationships. Agents may use a model, but only an isolated, deterministic policy checker holding the owner's signed mandate can cause a signature, and no model transforms data or grants access at run time. Data-plane calls use short-lived OAuth tokens bound to the agreement and to the caller's key, and each entity keeps its own tested, signed mapping to shared capability profiles. Discovery is a signed document at /.well-known/selo, optionally SVCB or a DAWN mechanism. The proposed IETF core is three pieces (the signed agreement object, the negotiation messages and state machine, an agreement-bound access token profile); trust frameworks, KYB, payments, marketplace search and individuals as parties are out of scope, no protocol is specified, and the author declares a pending patent application with an IPR disclosure to follow.
flowchart TB
subgraph EA["Entity A"]
A1["Agent, may use a model"] -->|"typed proposals"| P1["Policy checker: no model, key + mandate"]
P1 -->|"signs agreement"| R1["Runtime + gateway, pinned mapping"]
R1 --> API1["Platform API"]
end
subgraph EB["Entity B"]
A2["Agent, may use a model"] -->|"typed proposals"| P2["Policy checker: no model, key + mandate"]
P2 -->|"signs agreement"| R2["Runtime + gateway, pinned mapping"]
R2 --> API2["Platform API"]
end
A1 <-->|"control plane: verify, disclose, negotiate"| A2
R1 <-->|"data plane: agreement-bound tokens"| R2
S[("Issuers, trust list, issuance log, profile registries")] -.-> A1
S -.-> A2
Selo names DAWN as its discovery venue and WIMSE for cross-organization delegation, so it is a prospective consumer of his agent directory work and a test of whether a DAWN record can carry entity certificate chains and assurance levels. Its rule that no model can cause a signature is really an attestation requirement on the policy checker, which RATS could supply.
IETF 127: no WG slot; hallway topic. The draft announces a Side meeting Mon 16 Nov 17:30-19:00 PST, Continental 7; discussion on the dispatch list.
A verdict record is a canonical JSON object, bound to its issuer, that states a result reached on a claim, cites its evidence by SHA-256 digest (a SCITT Receipt, an Agent Action Capsule, or a record in any other format), and names the rule by which the result can be re-derived. Two kinds are defined: check-run for a mechanical rule, and model-decision for a decision by an AI model, where the rule name judgement declares that nothing can re-derive the result. A checker runs three steps (a check is present, every cited item is held and verifies under its format, the rule re-run gives the stated result) and reports one of four states: released, held, refused or no verdict. Two checkers with the same inputs must give the same answer, no step reads the checker's clock, and inputs must come from outside the issuer's reach. A checker must not print a word such as verified, valid, true or OK, because a check never shows that a claim is true, only that the issuer asserted it on the stated evidence under the stated rule. Records are versioned with supersession, and a section covers reassessment when evidence is withdrawn. Signing, time-stamping and transport are left to the container (a SCITT Signed Statement is the suggested one), no receipt format is defined or changed, and the draft positions itself as one way to record the Auditor's determination in the AUDIT architecture.
flowchart LR AG["Agent action"] --> CAP["Agent Action Capsule: disposition, effect binding"] CAP --> TS["SCITT Transparency Service"] TS --> RC["Receipt"] CAP -->|"cited by sha256"| VR["Verdict record: claim, evidence, rule, result"] RC -->|"cited by sha256"| VR VR -->|"carried as a SCITT Signed Statement"| TS CH["Checker, no access to issuer"] -->|"verifies evidence, re-runs rule"| VR CH --> ST["State: released, held, refused, no verdict"]
A SCITT statement profile that records one agent action with its disposition (executed, blocked, denied, errored, timed out), the constraints evaluated, a confirmed-effect binding that separates "dispatched" from "observed", and an honest human-in-the-loop flag; Ed25519 COSE_Sign1 producer envelopes sign a signer-independent Capsule ID, and a Capsule is recorded on every verdict, refusals included. SCITT Wed 18 Nov 16:30 PST.
This is the most concrete object proposed for the output side of the AUDIT BoF, and it will shape where an AUDIT charter draws the line between evidence and verdict. Citing evidence by digest with a named format is also the natural slot for RATS attestation results and SEAT evidence, which is where his attestation and audit interests meet.
model-decision with rule judgement only shows what a decision rested on: is that enough for the auditor use cases in the BoF?IETF 127: AUDIT BoF BoF Tue 17 Nov 09:00 PST; related Capsule at SCITT Wed 18 Nov 16:30 PST.
Defines two signed JWTs and how they travel on HTTP requests. An Agent Passport (agent-passport+jwt, at most 24 hours) binds an agent's holder key to an accountable controller origin and to a build: a model identifier, a JCS digest of the configuration (system instructions, tool allow-list, self-enforced policy), the tools with optional attestation digests, and a code digest; any change to the build requires a new passport. The controller lists its authorised issuers at /.well-known/agent-controller, and issuers publish keys through /.well-known/agent-passport-issuer. A Delegation Receipt (agent-delegation+jwt) records one step of authority from a principal, through a delegator, to a delegate key; chains can only attenuate, and the relying party checks each step against the one before and intersects the result with the principal's own access. Requests carry the passports and receipts in header fields covered by an RFC 9421 signature with the acting agent's key, so verification needs no per-request issuer contact, with status lists for revocation. Non-goals are explicit: the draft does not prove that the agent actually runs the named build (remote attestation is out of scope, though an issuer may require it), trust in controllers, issuers and builds is local policy, and reputation is excluded. It presents itself as complementary to the WIMSE and OAuth agent authentication profile and to AAuth, adding build binding, verifiable cross-organization attenuation and controller accountability.
sequenceDiagram participant AS as Principal AS participant I as Passport issuer participant A as Agent A participant B as Agent B participant RP as Relying party A->>I: holder key + build I-->>A: Passport A B->>I: holder key + build I-->>B: Passport B AS->>A: root Delegation Receipt bound to A key A->>B: child receipt, attenuated, signed with A key B->>RP: request signed by B with passports B, A and receipts RP->>I: issuer metadata and status lists, cached RP->>RP: check controller, chain narrows, intersect with principal RP-->>B: decision
Build-bound identity with attenuating cross-organization delegation sits exactly between WIMSE workload identity and OAuth token exchange, the gap his agent identity and delegation work is aimed at. Its admission that the build digest is only the issuer's assertion is an open invitation to make RATS or SEAT evidence an issuance condition.
act nesting and to WIMSE cross-org delegation; is a new token type really needed?IETF 127: Side meeting Tue 17 Nov 18:45-19:45 PST, Plaza A; WIMSE Wed 18 Nov 14:00 PST (AOB); joint WIMSE/OAuth Thu 19 Nov 14:00 PST. WIMSE Wed overlaps CoRE, which Jaime chairs: the side meeting and the Thursday joint session are the realistic slots.
Describes an intent-based system in which a User Agent sends a natural-language intent to a Cloud Agent or an Edge Agent, each holding an Intent Translator that turns it into a high-level policy and refines that into specific intents for Worker Agents on devices: routers and switches, firewalls and web filters, robots, software-defined vehicles, IoT devices, drones and urban air mobility. Agents talk to each other over A2A and reach their own modules and tools over MCP. A Knowledge Base holds policy knowledge (target schemas, valid examples, semantic constraints) and an Agent Catalog of A2A Agent Cards used to select Worker Agents. Results travel back up the chain as reports. The scope is managed domains (data centres, enterprise, campus and home networks), not the public Internet. Agent discovery is left out of scope, with pointers to DAWN drafts for reactive DNS and mDNS discovery and to I2NSF registration for proactive registration. Security is three paragraphs (mutual authentication, logging of harmful intents, an independent auditor), no data models are defined yet, and the substantive content is about six pages.
flowchart TB
U["User Agent: smartphone, computer"] -->|"intent, A2A"| C["Cloud Agent + Intent Translator"]
U -->|"intent, A2A"| E["Edge Agent + Intent Translator"]
C <-->|"A2A"| E
C --- KB1[("Knowledge Base: policy knowledge + Agent Catalog")]
E --- KB2[("Knowledge Base: policy knowledge + Agent Catalog")]
E -->|"specific intent"| W1["Worker Agent: router, switch"]
E -->|"specific intent"| W2["Worker Agent: firewall, web filter"]
E -->|"specific intent"| W3["Worker Agent: robot, SDV, IoT"]
W1 -->|"MCP"| M1["Modules"]
W2 -->|"MCP"| M2["Modules"]
W3 -->|"MCP"| M3["Modules"]
W3 -.->|"reports"| E
E -.->|"reports"| C
The framework rests on an Agent Catalog of Agent Cards and defers discovery to DAWN, which makes it a concrete research use case for Agent Resource Discovery in managed networks, including constrained IoT Worker Agents where CoRE resource directories already exist. The thin security section leaves room for workload identity and attestation of Worker Agents before they are trusted with intents.
IETF 127: NMRG Mon 16 Nov 11:30 PST and Tue 17 Nov 11:00 PST. Mon 11:30 overlaps AIPREF.
The profile draft was renamed to draft-skyfire-oauth-kyapay-token; its -03 (2 Nov 2026, 46 pages) is the current text, so the summary below describes that version.
Defines JWT formats for agent identity (KYA, "Know Your Agent"), payment (PAY) and combined KYA-PAY tokens, with media types application/kya+jwt, pay+jwt and kya-pay+jwt. A KYA token carries layered identity: hid for the human or organizational principal (with verifier and verified flags), apd for the agent platform, aid for the agent instance, plus scope, audience and target identification; PAY tokens add amount, currency, settlement information and card settlement scope. Token issuers are framed as the client-side counterpart of certificate authorities, publish keys at a well-known location and sign with ES256. Tokens are bearer by default; a cnf claim can sender-constrain them, with the HTTP Message Signature check specified in the "using" companion. The stated aim includes letting agents "bypass common blocking mechanisms" by giving web security systems a strong signal that a human authorised the request. The protocols that use the tokens are explicitly not defined, and the payment token could be split out if a WG adopts the work.
sequenceDiagram participant P as Human principal participant AG as Agent on platform participant IS as KYAPay issuer participant SI as Security intermediary participant AS as Authorization server participant RS as Resource or MCP server P->>IS: identity verified, ivm, amr, aml recorded AG->>IS: request KYA or PAY token IS-->>AG: signed JWT with hid, apd, aid, scope, optional cnf AG->>SI: request with KYAPay-Token header SI->>SI: validate with issuer keys, allow, step-up or block AG->>RS: fetch protected resource metadata, RFC 9728 AG->>AS: RFC 8693 token exchange, subject_token_type kyapay AS-->>AG: OAuth access token, bearer or DPoP AG->>RS: call with access token
KYAPay-Token header; bot managers, fraud and ATO systems validate it and decide allow, step-up or block.kyapay:iss|ori|apd|aid or with a CIMD client, plus a kyapay grant profile and kyapay_issuers_supported server metadata.KYAPay-Token header and Accept-KYAPay-Token response, asking "did a verified human authorise this agent" rather than "is this a human", with proof of possession over HTTP Message Signatures as the planned evolution.ivm claim and IANA registry of identity verification methods, from database PII checks and ID documents to video and in-person, with business registry, TIN, business document, beneficial owner and representative checks added in -03.amr values app, bg, sna (silent network authentication, citing CAMARA Number Verification), email, code, url, push, facliv, passkey, mdc and uv, a redefined tel, and guidance on composing channel and challenge values.aml claim and registry with ofac, sanc, watch, pep and adv screening values, normatively referencing vendor web pages for their definitions.KYAPay is a principal, platform and agent identity stack expressed as OAuth tokens, so it competes for the same space as agent identity and delegation in WIMSE and OAuth and as the passport draft's controller and build binding. The sna value brings mobile network number verification into the OAuth AMR registry, a direct operator-network hook in agent authentication.
cnf plus HTTP Message Signatures, aligned with Web Bot Auth, before any adoption call?act chains; can a multi-hop delegation be expressed, or only one agent?IETF 127: OAuth Wed 18 Nov 09:00 PST and Fri 20 Nov 09:00 PST; agent items may land in the joint WIMSE/OAuth session Thu 19 Nov 14:00 PST. Wed 09:00 overlaps GEN area.
| # | Read | Why in this position |
|---|---|---|
| 1 | draft-le-agent-passport-00 (25 pp) | Closest to his agent identity and delegation work; the Tue evening side meeting is where -01 open issues get shaped, and WIMSE Wed clashes with CoRE. |
| 2 | draft-wright-verdict-record-02 (24 pp), skim the Capsule | AUDIT BoF is Tue 09:00, the first of his agent sessions where a concrete output format is on the table. |
| 3 | Web Bot Auth group, protocol -01 then registry -04 | The Signature Agent Card is the nearest thing to a DAWN agent directory record, and the -01 type removal is the question to raise Tue 16:30. |
| 4 | kyapay-token-03 and token-exchange-02 (skip profile-02) | OAuth Wed and the Thu joint session; competes with WIMSE agent identity, and the bearer default is worth challenging. |
| 5 | draft-hamlett-selo-architecture-00 (30 pp) | Names DAWN as its discovery venue and WIMSE for delegation; architecture only, so Sections 1, 5 and 6 are enough before the Mon side meeting. |
| 6 | draft-jeong-nmrg-agentic-ai-framework-00 (17 pp) | Short, a useful DAWN and constrained-device use case, but research group and overlapping AIPREF on Monday. |
| 7 | draft-vaughan-machine-readability-01 (19 pp) | Good vocabulary for reviewing discovery records, but conceptual and furthest from his own drafts. |
Generated 6 November 2026 from the public draft text. Times as listed on the personal agenda.