What happened

On 10 November 2026, at KubeCon + CloudNativeCon North America in Salt Lake City, the Linux Foundation said the Agent Name Service (ANS) is now one of its hosted projects. The foundation first announced its intent to launch the project in June 2026. ANS is an open standard for giving AI agents names, then letting other software find them and check they're genuine. It's built on the Domain Name System, which most organisations already run.

A technical steering committee under the foundation will govern the specification from now on. The foundation says eleven organisations have contributed to or adopted the standard since June: Cisco, Cloudflare, DigiCert, DistributedApps.ai, GoDaddy, HOL, Infoblox, Radix, Salesforce, OWASP and Trustmodel.ai. Ken Huang, CEO of DistributedApps.ai, led the early development work and will sit on the steering committee. DigiCert's Brian Trzupek summed up the idea as "a passport for AI agents."

Two integrations already exist. AGNTCY's Agent Directory Service, which is also a Linux Foundation project, can verify ANS identities. A tool called ANS Finder implements the Agentic Resource Discovery specification and passes its registry conformance tests.

Why agents need names

Agents increasingly call other agents, APIs and Model Context Protocol servers on someone's behalf. When a request arrives, the receiving system wants to know who runs the caller, which build of the software it is, and whether that registration is still valid. Today most teams answer those questions with API keys and private allowlists, and they don't carry over from one platform to the next.

ANS answers them with tools the internet already trusts. It ties an agent to the domain its operator controls, to standard TLS certificates, and to a public, append-only log of registration events. The specification doesn't add a new root of trust, which makes it easier to adopt with existing infrastructure.

How it works

Every registration gets a versioned name in the form ans://v1.0.0.payments.example.com. The version uses plain major.minor.patch numbers, and each new version counts as a separate registration with its own status. So a buggy release can be pulled without touching the others.

Three roles do the work:

  • The operator sends the agent's hostname, version, endpoints and certificate details to a Registration Authority (RA).
  • The RA checks domain control using ACME-style DNS-01 or HTTP-01 challenges, the same ones used to issue free web certificates. It can also issue an Identity Certificate from its own private certificate authority, with the ANS name embedded so the agent can use it to authenticate as a client.
  • The Transparency Log is a Merkle-tree log modelled on Certificate Transparency. It seals each signed lifecycle event before an agent goes live, and it gives out receipts, short-lived status tokens and signed checkpoints. Outsiders can use these to spot it if anyone rewrites history.

A registration moves from pending validation, to pending DNS, to active. Later it can be marked deprecated or revoked.

On the DNS side, the operator publishes records the RA supplies. The default discovery format is an SVCB record at the agent's hostname. There's also an opt-in TXT format, which version 1 used. An _ans-badge TXT record is required, and it points at the agent's entry in the log. Operators can also add a TLSA record for DANE, though verifiers will only trust it when DNSSEC validates it.

How a caller is checked

When one agent contacts another, the receiving side checks three things. Is the certificate tied to this agent in the log? Is the registration still current? Does the caller actually hold the private key? The project's own documentation stresses that the third check matters: public badges and receipts can't, on their own, prove who's calling.

Verifiers pick one of two tiers. The badge tier asks the log about the agent live, at connection time. The SCITT tier checks signed receipts and status tokens locally, so it keeps working when the log is slow or out of reach. They also pick how the caller proves it holds the key. Option A is mutual TLS with the Identity Certificate. Option B is DPoP (RFC 9449): the caller sends a signed, single-use proof in an HTTP header that's bound to the request. Option B is useful when a gateway or load balancer ends the TLS connection before traffic reaches the agent.

The standards picture

There are two layers to keep track of. The project's GitHub specification repository holds seven layered documents, numbered ANS-0 to ANS-6, covering identity anchoring, registration, naming, DNS publication, transparency, integrity monitoring and agent-to-agent authentication. Underneath sits an IETF Internet-Draft, draft-narajala-courtney-ansv2, now at revision 02 (26 October 2026). Its authors come from GoDaddy, OWASP, DistributedApps.ai and Cisco. It's an independent submission with Informational status, so it has no formal standing in the IETF standards process, and it expires on 29 April 2027 unless it's updated.

Limits worth knowing

  • Maturity. The ANS website describes the published specifications as drafts still under active development. Neither the spec repository nor the reference implementation had any tagged releases when we checked.
  • Revocation lag. When the RA revokes an agent, it writes a revocation event to the log. On the SCITT tier, the IETF draft puts the worst-case delay at about two hours and ten minutes at default settings. That's a one-hour token lifetime, up to ten minutes of clock skew, and a grace window of up to one more token lifetime. Teams with tight risk budgets should plan around that.
  • Scope. ANS proves who an agent is. It doesn't decide what the agent may do, delegate a user's permissions, or prove what code a remote machine is running. Trust scoring lives in a separate project. Your application still has to handle authorisation and policy.
  • Reference code gaps. The Go implementation's default DNS verifier is a no-op that accepts any DNS state, and the bundled DNS server is for development only. Real ACME DNS-01 wiring, plus adapters for cloud key management, Postgres and managed DNS, are listed as follow-up work for contributors.

Costs and licensing

None of the official pages we read mention fees. Running the reference RA and log yourself costs only compute. Commercial registration authorities may charge, so check before you commit. The specification repository is under Apache 2.0. In the implementation, new code is Apache 2.0 and some older contributions keep MIT terms, so a distribution that includes both has to meet both licences.

What builders can do now

  • Run it locally. The reference repo builds with Go 1.26 or later. It ships an RA, a log, an offline ans-verify tool and a demo script that registers an agent and checks its receipt from start to finish. Docker Compose works too.
  • Add verification. The project lists SDKs in Go, Rust and Java for adding checks to existing services.
  • Use what your platform offers. MuleSoft's API Manager can already register agents with GoDaddy's ANS. It needs a custom domain (MuleSoft-managed domains such as cloudhub.io aren't supported), and the registered hostname can't be changed afterwards.
  • Check your DNS. New Zealand teams can anchor agents to .nz or .co.nz names. Before you do, we'd confirm your DNS host supports SVCB records and DNSSEC.

Our view: it's worth prototyping now, so you're ready when tagged releases and a firmer revocation story arrive. We'd hold off on production use until then.

Sources

Build with people working on the same things

Join the Auckland Society for AI: NZ$599 a year with a ForgeBox and the members' projects area. Our Auckland meetups are open to everyone.

Become a member