Skip to content
IAM Insights

Explainers · Glossary

ANS

ANS v2 · Agent Name Service · Agent Naming Service · draft-narajala-ans · draft-narajala-courtney-ansv2

By Boston Identity · Explained · 6 min read

The short answer

ANS is a proposal to name an AI agent by a software version plus a domain its owner controls, so another agent can check, through certificates, DNS records and a tamper-evident log, that it is really that agent.

What it is
A protocol
Defined by
Individual IETF Internet-Draft (-01), now expired
Who it's for
Teams whose agents deal with other companies' agents
In our briefings
20By month of the source: 2025-12: 1, 2026-01: 0, 2026-02: 0, 2026-03: 1, 2026-04: 1, 2026-05: 0, 2026-06: 3, 2026-07: 2, 2026-08: 0, 2026-09: 2, 2026-10: 3, 2026-11: 00by month, last 12
01

What is ANS?

The Agent Name Service (ANS) is a proposal for giving each AI agent a name that another agent can check before trusting it. Version two names an agent by a software version in front of a domain name, as in ans://v1.0.0.sentiment-analyzer.example.com.§5.1 The domain stays fixed for the life of the agent, and the version number changes whenever its software changes.§1.5

The authors start from a payment agent told to wire $50,000 to a supplier's agent. A look-alike domain with a valid certificate would fool it, and so would a real supplier that quietly changed the audited code behind its endpoint.§1.1 The agent protocols MCP (Model Context Protocol) and A2A (Agent2Agent) say how agents talk, but they leave open who the agent is and whether to trust it.§1

ANS v2 is a draft, not a standard. It is an individual submission to the IETF (Internet Engineering Task Force), published on 13 April 2026 with an intended status of informational, and the IETF's datatracker now lists it as expired.S1

ANS is not a search directory for finding an agent that can do a job. The first version leaned that way, but version two leaves discovery to independent services and only answers whether a named agent is who it says.1,2

02

How does ANS work?

The hosting platform runs the agent's code at its domain. The registration authority (RA) checks and records; two certificate authorities (CAs) sign certificates; a transparency log keeps an append-only record; and a caller, another agent, does the checking.§4

Figure steps
  1. The hosting platform asks to register
  2. The RA checks domain control
  3. The RA obtains two certificates
  4. The RA seals the event in the log
  5. The hosting platform publishes DNS records
  6. The caller checks before it connects
Fig. 1

The hosting platform registers the agent with the registration authority, which checks the domain, gets two certificates and seals the event in the log; a caller then checks as much of that as it chooses.

Drawn from the ANS v2 draft, sections 6.1 and 7.1

Swipe for the full figure →

  1. The hosting platform asks to register

    The hosting platform sends the RA the agent's domain, version and display name, at least one endpoint, and a certificate signing request (CSR) for an identity certificate.§7.1.1,§5.1 The draft keeps that minimum low on purpose. A Legal Entity Identifier (LEI) and metadata taken from the agent's A2A or MCP card are optional.§1.6

  2. The RA checks domain control

    The RA confirms that the requester controls the domain with ACME (Automatic Certificate Management Environment), using the DNS-01 or HTTP-01 challenge.§7.1.2 It also fetches each metadata URL and compares its hash. It checks for DNSSEC (signed DNS) as well, but only warns if it is missing.§7.1.2

  3. The RA obtains two certificates

    A public CA issues a server certificate for the stable domain, used for ordinary TLS (Transport Layer Security).§4.5 A private CA issues an identity certificate that carries the versioned ANS name. A public CA cannot issue that one, because industry rules for public certificates ban the kind of name field it needs.§5.4

  4. The RA seals the event in the log

    The RA submits a signed registration event to the transparency log, which seals it into an append-only structure.§4.3 The draft calls this the point of no return. After sealing, the identity fields cannot change, and any change needs a new version and a new log event.§7

  5. The hosting platform publishes DNS records

    The RA hands back the certificates and the content of the DNS records, and the hosting platform publishes them.§7.1.2 An _ans record says which protocol the agent speaks and where its metadata lives. An _ans-badge record points to the agent's entry in the log, and a TLSA record carries the server certificate's fingerprint.§7.4

  6. The caller checks before it connects

    A caller that validates only the certificate chain reaches the bronze tier. Matching the certificate's fingerprint to DNS through DANE (DNS-based authentication, which needs DNSSEC) reaches silver, and confirming the entry in the log reaches gold.§6.1 The tier describes what this caller checked, not a grade the RA assigned.§6.1 Between two registered agents, the caller also presents its own identity certificate over mutual TLS.§6.4

03

Why does ANS matter now?

Agent naming has split into two camps this year, and ANS moved between them. Its first version, a May 2025 specification for the OWASP Agentic Security Initiative, packed protocol, capability, provider and version into a six-part name.3 In April 2026 the authors rebuilt it on DNS, with an authority that checks domain control the way certificate issuers do.2 That puts ANS opposite the Decentralized Identity Foundation's KYA-OS, which names agents only by decentralized identifiers (DIDs).4 The gap is narrower than it looks, because the first DID methods the foundation recommends both resolve from web domains.5

Nobody runs it yet. The draft has no standards body behind it, adds an author from the domain registrar GoDaddy, and names no operator.6 Who allocates names, what registration costs and who holds the root certificate are left to a future governance paper.6 The same hole runs across the field: at least five individual drafts on how agents describe themselves each describe a registry or an authority, and none says who runs it.7

04

Where does it fit?

Compared withHow ANS differs
KYA-OSKYA-OS names an agent only by a decentralized identifier, and the agent generates its own key, which no registry may hold. ANS anchors the name to a domain that a registration authority has checked.4,2
A2AA2A says how agents negotiate and hand off tasks, and leaves open who the agent is. ANS answers that question, and its _ans record can point a caller to the agent's A2A card.§1,§7.4.1
Agent registryThe ANS draft does not mention any platform's own agent registry. How a domain-anchored name relates to the identity a company already gives its agents and workloads is left open.1
05

What goes wrong?

  • Collusion breaks it. The authors say their detection property does not hold if the RA and log operators collude, and the same entity may run both for a given registration.§9.2,§12.8 Their remedies are checkpoints anchored to a separate ledger and federation across operators, and a multi-RA federation is listed as future work.§11
  • A domain is not a company. The RA checks control of a domain, and the draft's own scenario notes that a domain-validated certificate proves only that someone controls the domain.§1.1 Binding the agent to a legal entity, by an LEI or a DID, is optional.§1.6
  • Checks step down. Without DNSSEC, the client must treat the DANE record as unusable and falls back to ordinary certificate checks.§6.1 If the log is unreachable, the caller drops to certificates plus DANE and records the downgrade.§6.4
  • A version number is a declaration. The draft lists runtime checks of the running code, through zero-knowledge proofs or trusted execution environments, as future work.§11
  • The trust root is private. Identity certificates come from a private CA, so each verifier has to install its root, and the SDK installs the bootstrapping RA's root until a federation exists.§4.8,§8.4
  • The rules of the market are unwritten: name allocation, fees, dispute settlement and stewardship of the root CA wait for a governance white paper.§11
06

What should a security leader ask?

  1. 01

    Which outside agents do ours accept requests from today, and what do we check about their identity before acting?

  2. 02

    If a vendor offers ANS-style verification, who runs the registration authority and the log, and are they different operators?

  3. 03

    Do the domains our agents run on have DNSSEC, which the silver tier needs?

  4. 04

    When an agent's code changes, does our pipeline register a new version, or does the old certificate keep vouching for it?

  5. 05

    How would a domain-based agent name map to the workload identities we already issue?

07

Sources

›Evidence: 7 claims from 5 briefings
  1. 1BI-FIND-2026-0305 · The Agent Name Service draft was rebuilt on DNS, with a GoDaddy author, and no standards body has adopted itVersion two answers whether a named agent is who it says, and hands the search for an agent to others. The registration authority does not run discovery: 'agent discovery flows from the Event Stream through independent indexing services', and how a caller searches by capability is not specified. The draft positions itself as the trust layer over a separate discovery proposal, DNS-AID, and over A2A and MCP, which it says 'explicitly defer the question of who the agent is'. It does not mention SPIFFE, the IETF's workload-identity work or any platform's own agent registry, so how a domain-anchored name relates to the identity an enterprise already gives its workloads is left unaddressed.
  2. 2BI-FIND-2026-0302 · The Agent Name Service draft was rebuilt on DNS, with a GoDaddy author, and no standards body has adopted itThe Agent Name Service has been rebuilt on the domain name system. In version two an agent's name is a software version in front of a domain its publisher controls, in the form ans://v1.0.0.agent.example.com. A registration authority checks control of the domain the way certificate issuers do, then issues two certificates, one public for the server and one private that carries the agent name. Callers find the agent through a DNS record that says which protocol it speaks, A2A, MCP or plain HTTP, and where its description lives. Every code change needs a new version and a new registration. The first draft's design, a directory that resolved a capability to a decentralized identifier, is gone.
  3. 3BI-FIND-2025-0435 · The first Agent Name Service paper puts capabilities in the name and leaves protocol translation and governance outIn their May 2025 specification for the OWASP Agentic Security Initiative, Ken Huang, Vineeth Sai Narajala, Idan Habler, and Akram Sheriff encode an agent’s protocol, capability, provider, and version directly into a six-part name: Protocol://AgentID.agentCapability.Provider.vVersion.Extension, such as a2a://textProcessor.DocumentTranslation.AcmeCorp.v2.1.hipaa. A Registration Authority checks the registrant’s identity, obtains an X.509 certificate from a Certificate Authority for the agent’s Certificate Signing Request, and stores the record in the Agent Registry; decentralized identifiers (DIDs) appear only once in passing as a database field and never in the schemas or resolution algorithm. At lookup, the resolution algorithm selects the highest compatible version that "MUST adhere to Semantic Versioning standards" (ranking pre-release tags such as -rc1 and -beta below stable versions), verifies the registry’s signature and CA chain via CRL or OCSP, and returns an endpoint record where "the Agent Registry MUST include a Time-To-Live (TTL) value with each resolved Endpoint," with a recommended default of 300 seconds (5 minutes) after which resolvers must revalidate.
  4. 4BI-FIND-2026-0640 · DIF's agent spec names every AI agent by a DID; the IETF's Agent Name Service switched to domain namesKYA-OS names AI agents by decentralized identifiers and nothing else. Implementations must support did:key and did:web; no DNS name, OAuth client id or other identifier counts as an agent's identity, and the agent generates its own key, which no registry or verifier may ever hold. That is narrower than DIF described it in April 2026, when the first conformance level was to use existing OIDC and JWT identifiers; the v1.0 conformance document's Level 1 begins by deriving a did:key. It puts the Decentralized Identity Foundation's agent spec on the opposite side from the IETF's Agent Name Service, which moved from DIDs to domain names in its April 2026 revision.
  5. 5BI-FIND-2026-0654 · The first DID methods DIF and Trust over IP recommend are both web-based; blockchain ones are still in reviewThe first DID methods to win the Decentralized Identity Foundation and Trust over IP's recommendation are both web-based: did:webvh and did:webplus passed in June 2026, while the blockchain methods did:ethr and did:cid were still in their 60-day review and did:peer had not yet been submitted. DIF says it aims to approve web-based, decentralized and ephemeral methods, and that a recommendation means proven production use, not superiority. For agent naming, it means the DID methods with the most standing today resolve from web domains, which narrows the gap between the DID camp and the IETF's domain-based Agent Name Service.
  6. 6BI-FIND-2026-0304 · The Agent Name Service draft was rebuilt on DNS, with a GoDaddy author, and no standards body has adopted itEleven months after the first draft, the naming proposal still has no standards body behind it and nobody running it. The replacement is an individual submission the IETF has not adopted, set to expire on 15 October 2026. It adds an author from GoDaddy, a domain registrar, and names no operator. Who allocates names, what registration costs, how disputes are settled and who holds the root certificate are all assigned to a governance paper listed under future work, and running more than one registration authority is future work too. Any organization may run one, the draft says. None is named as doing so.
  7. 7BI-FIND-2026-0457 · Cisco's SD-Card puts an A2A Agent Card in a selective-disclosure token, and leaves its registry out of scopeThe draft expired on 1 September 2026 at revision 02 with no working group, no implementation and no list discussion we could find. The A2A specification it extends chose a different answer in the meantime: the current A2A text signs an Agent Card as a JWS over a canonical JSON form (AgentCardSignature) and does not mention selective disclosure. On the datatracker, agent self-description now has at least five individual drafts and no working group: this one, a framework-neutral AgentCard (April 2026), an OpenA2A Agent Identity Protocol at revision 04 (October 2026), an A2A trust model at revision 03 (September) and a WebFinger profile for naming agents (August). Each describes a registry or an authority; none says who runs it.

›In our briefings

The claim each of the newest briefings on ANS would have you leave with.

  1. In AGTP -09 the Agent-ID header on a request is self-asserted: the server records what the client wrote. The base check, Level 2, resolves that ID to an Agent Identity Document signed by the issuing governance platform and confirms the ID matches the hash of the agent's Agent Genesis record. Binding the ID and its Authority-Scope to the TLS connection, Level 3, takes the Agent Certificate extension (AGTP-CERT). The draft marks that extension optional, and Hood lists it, with Agent Genesis, among mechanisms that may be subject to his pending patents; section 11.7 says implementers of the core specification are not affected by claims on those extensions and formats, and Hood offers a royalty-free licence. No step in the Level 2 procedure has the caller prove it holds a key, so we read Level 2 as showing that an ID exists and is properly signed, not that the caller is that agent. An OAuth, OIDC or SPIFFE credential can ride in the Authorization header, but the draft says that credential identifies the principal the agent acts for. The threat model itself says full mitigation of Agent Genesis spoofing requires AGTP-CERT.

    AGTP checks that an agent ID is signed; binding it to the connection is an add-on in the author's patent noticeBI-FIND-2026-1898

  2. Drake's problem statement asks for an agent identity, which he calls 'identity0', that survives restarts, redeployment, software updates, replaced hardware anchors and a change of operator, and that cannot be sold or handed to a different entity. Who runs the agent is 'an authorization-layer relationship, not a component of the agent's identity'. Against Sybil attacks, the strongest tier makes each extra identity consume a scarce resource, which the draft says is not a fee. Revision -01 widens the free requirement: enrolling, keeping, authenticating with and verifying a base identity cost the actor nothing, while content-neutral rate, volume and abuse limits stay allowed. The draft says it is not a survey. It cites three other agent-identity drafts, and neither revision's text mentions OAuth, SPIFFE, X.509 or Agent Name Service, so in our reading its case that existing work leaves a gap rests on those three drafts. On 8 October the author offered the series to the IETF's WIMSE workload identity group as its foundation; his 1 October request for guidance on the DISPATCH list, where the IETF sorts new work, had no reply when we checked on 28 October.

    Drake's revised draft: an agent identity outlives its operator but is never reissued once its last key is goneBI-FIND-2026-0753

  3. Revision -04 of the IETF's workload identifier draft turns the question of whether two workload names are the same into fixed steps. Revision -03 already required comparing the complete URI; -04 says how. Implementations MUST validate the identifier as received, before normalization can remove evidence of prohibited syntax; a '.' or '..' path segment is rejected even when percent-encoded. An internationalized trust domain MUST be written as A-labels. Consumers then apply RFC 3986's syntax-based normalization, decoding percent-encoded unreserved characters no more than once and leaving an encoded slash (%2F) encoded, and compare the complete normalized URIs byte-for-byte. Names with different schemes are distinct, the same rules decide uniqueness and authorization, and trust-domain allowlists MUST match on the ASCII form. The draft's stated reason is that "inconsistent processing can allow authorization checks to be bypassed". These rules are new in the -04 text, but the revision's own change log (Appendix B.4) does not list them: it names editorial fixes, stale references, an updated mapping section and uniqueness clarifications, and leaves out the loosened IP-address rule too.

    The IETF names workloads with a URI on the owner's domain; revision -04 pins down how two names are comparedBI-FIND-2026-1524

›Cited in 20 briefings

Every briefing that discusses this term, newest first.

  1. AGTP checks that an agent ID is signed; binding it to the connection is an add-on in the author's patent notice

    Draft · · IETF Datatracker · Chris Hood, Nomotic, Inc. · Jun 28, 2026

  2. Drake's revised draft: an agent identity outlives its operator but is never reissued once its last key is gone

    Draft · · IETF Datatracker · Christopher Drake, 1id.com · Sep 25, 2026

  3. The IETF names workloads with a URI on the owner's domain; revision -04 pins down how two names are compared

    Draft · · IETF Datatracker · Yaroslav Rosomakho et al. · Oct 23, 2026

  4. Agents and MCP servers get broad standing access because it is quick or easy, three voices say, and three vendors now pitch a gateway in front of MCP

    Field signals · · 9 voices from the field · Sep 2025 to Oct 2026

  5. At OWASP, one poisoned memory entry made a demo agent send a Q4 report to an attacker server

    Video · · OWASP Foundation · Recorded Jun 2026 · Uploaded October 21, 2026

  6. A post-quantum agent certificate draft warns that its own declared scopes are not proof of behaviour

    Draft · · IETF Datatracker · Reuben Burls · Jul 21, 2026

  7. Okta's expired SCIM agent draft recorded certificates and OAuth clients; the joint draft after it dropped both

    Draft · · IETF Datatracker · Macy Abbey et al. · Oct 16, 2025

  8. A decentralized ID design for agents proves who is liable but can collect only from collateral locked in advance

    Paper · · arXiv · Ahammod Munim Ramil et al. · Oct 12, 2026

  9. DIF's agent spec names every AI agent by a DID; the IETF's Agent Name Service switched to domain names

    Article · · Decentralized Identity Foundation (KYA-OS) · Dylan Hobbs et al. · Jul 29, 2026

  10. The first DID methods DIF and Trust over IP recommend are both web-based; blockchain ones are still in review

    Article · · Decentralized Identity Foundation · Decentralized Identity Foundation et al. · Jun 19, 2026

  11. In Accenture's agent-delegation design, no outside party checks the user, and revocation is left open

    Paper · · arXiv · Srikumar K. Subramanian et al. · Sep 25, 2026

  12. Cisco's SD-Card puts an A2A Agent Card in a selective-disclosure token, and leaves its registry out of scope

    Draft · · IETF Datatracker · Suhas Nandakumar et al. · Mar 2, 2026

  13. The Agent Name Service draft was rebuilt on DNS, with a GoDaddy author, and no standards body has adopted it

    Draft · · IETF Datatracker · Scott Courtney et al. · Apr 13, 2026

  14. A Shenzhen paper ties each AI agent to a verified owner and its exact code, at 15 to 40 seconds of proving a step

    Paper · · arXiv · Zibin Lin et al. · Dec 19, 2025

  15. AAGATE maps NIST's AI risk framework onto Kubernetes, and its paper never times the 'millisecond kill switch'

    Paper · · arXiv · Ken Huang et al. · Oct 29, 2025

  16. An October SCIM draft gives agents their own records, with security still unfinished

    Draft · · IETF Datatracker · Macy Abbey et al. · Oct 11, 2025

  17. A 'provable' bound on agent-memory injection assumes independent layers and fills in none of its detection rates

    Paper · · arXiv · Ken Huang et al. · Aug 17, 2025

  18. A Cloud Security Alliance paper wants AI agents on a new identity stack, with OAuth kept for simple cases

    Paper · · arXiv (cs.CR preprint 2505.19301) / Cloud Security Alliance · Ken Huang et al. · May 25, 2025

  19. The first Agent Name Service paper puts capabilities in the name and leaves protocol translation and governance out

    Paper · · arXiv (cs.CR preprint 2505.10609) / OWASP Agentic Security Initiative · Ken Huang et al. · May 15, 2025

  20. MAESTRO names your AI agents' identity threats and leaves designing the controls to you

    Article · · Cloud Security Alliance · Ken Huang · Feb 6, 2025

In our columns

  1. A signed AI agent name is an IP address with a key, not a login

    Column · Oct 30, 2026

  2. However an AI agent is named, the trust behind it runs through a web domain

    Column · Oct 19, 2026

Related terms

What appears alongside ANS in the same briefings.

AgentDNSAgent registryAgent Communication Protocol (IBM)MAESTROIBMA2AGoDaddyAgent Network Protocol (ANP)

Was this explainer useful?

Share this explainer

Pass it on. Every claim has a permanent number, so what you share today still points at the same words next year.