Most agent-identity work assumes the agent has an organisational home: a domain, a certificate, a directory entry somebody maintains. MIT Media Lab, Cisco, Google and Project NANDA have published a design for the agents that have none — and been unusually explicit that finding an agent is not the same as trusting one.
What happened
On 24 October 2026, MIT Media Lab published “Making AI Agents Discoverable Without Owning a Domain”, co-authored by Ramesh Raskar and Pradyumna Chari (MIT Media Lab), Luca Muscariello and Vijoy Pandey (Cisco), Samuel Sharaf and Junjie Bu (Google) and Karan Bharadwaj (Project NANDA). It describes NandaIndex, an open resolution layer for what the authors call the agentic web.
The design is available as a request-for-comments technical paper, A Global Switchboard for the Agentic Web (v0.1), alongside a live site at nandaindex.org and a reference implementation. The publication is not a product launch: it is a design plus a set of supporting statements from companies including Ericsson, TCS, HCLTech, Wipro, Cognizant, Kyndryl, Nutanix, BigCommerce, WorkOS, Inkbox and Hello Haven. The article notes that naming a product does not imply that its provider supports the index.
The problem it addresses is narrow and concrete. Enterprise agent discovery starts from infrastructure the organisation controls: a published AI Catalog, DNS-based service discovery, a governed gateway. A neighbourhood bakery, a tutoring business or an individual with an ordinary email account has none of that. Their agent card may be hosted by one provider and their runtime by another. If two such agents need to exchange a task, there is no bilateral integration to fall back on.
How it works
NandaIndex resolves identifiers to pointers and stops there. Given a stable identifier — a domain, a provider-verified account, a DID, a resource identifier — it returns a typed discovery record in the AI Catalog entry format: an identifier, a media type, and either a URL or inline data. That record can point at a hosted A2A Agent Card, an AI Catalog, an Agentic Resource Discovery (ARD) directory, a DNS-based discovery path, or another native mechanism.
The authors are careful about what the index is not. It does not host the agent, does not decide whether an agent is trustworthy, and does not replace the protocols used to talk to it. Resolution is one step; the requester then follows the pointer, retrieves the agent card or directory, and invokes the resource under whatever authorisation that service requires.
Three deployment cases do most of the explanatory work:
- An enterprise with a published AI Catalog is resolvable through its own domain, and the index is optional — extra visibility or a fallback path, not a required intermediary.
- A small business that owns a domain but has no one to configure DNS records can keep its public identity under its own domain, host its agent card with one service and run its agent on a third, with a single index record joining the identifier to the card.
- An individual with only an email account is handled by anchoring the resource identifier to an accountable publisher domain — the agent-card host — and associating the person’s verified email with that record as a subject account. The email is the human-readable lookup key, not a claim that the individual controls the mail provider’s DNS zone.
Registration paths documented on the site include email confirmation for account-backed personal agents and DNS verification for domain-based records.
The boundary the authors draw
The passage that matters most for anyone doing KYA is the one that separates being listed from being allowed to act:
“A listing permits discovery; it does not authorize a requester to purchase goods, access personal information, or instruct an agent to act. Authentication, consent, spending limits, and audit mechanisms remain essential at the invocation layer.”
That is the correct division of labour, and it is also the sentence that keeps the announcement from being oversold. A resolution hit tells you where to look. It does not tell you who is accountable.
What the design leaves open
The article ends with an explicit list of unsolved problems, and it reads like a verification checklist:
- Identity binding. How does a resolver verify that a person controls the email account, or that a business controls the domain, associated with an agent — while preventing impersonation and name squatting? In the domainless case, the relationship between the account subject and the domain-anchored publisher identifier must be established cryptographically or otherwise verifiably. The authors do not claim it is solved.
- Trust and provenance. Discovery records should support authenticated publication, key rotation, revocation and verifiable agent-card references. A valid record is not proof that an agent is safe or capable.
- Federation. Independent resolvers need specified handoffs, consistent record-type interpretation, caching and freshness rules, and a named party responsible for inaccurate or disputed entries.
- Light weight. Registration and portability must not cost what an enterprise integration costs, or permissionless participation is theoretical.
Why it matters for KYA
Discovery sits upstream of verification, and it is not a substitute for it. You cannot verify an agent you cannot find, so a resolution layer is genuine infrastructure. But every question KYA exists to answer — who is the principal, what are the agent’s credentials, what may it do, what did it do — is downstream of the lookup. Two of the companies quoted make the same point in their own words: WorkOS’s Michael Grinich says “finding an AI agent and trusting it are two different problems”, and Inkbox’s Ray Liao frames the requirement as verifiable identity plus domain-name-like discoverability.
The domainless case forces the accountability question into the open. When an agent’s principal has no domain, somebody else’s domain ends up on the record. The NandaIndex answer — an accountable publisher domain plus a subject account — is a policy choice about who carries responsibility for the listing, not a cryptographic proof. That is worth naming plainly, because the same structure is being assembled elsewhere: hosted agent identities, agent registries with platform-issued handles, and enterprise identity providers that will happily issue credentials for an agent whose human owner they have never verified.
It makes the scale of the problem visible. The article points at citizen-facing agent projects such as Doot and KumbhDoot, where requiring every citizen to operate a domain and an agent catalog is not a realistic prerequisite. If personal agents reach public services at that scale, account-bound agent identity becomes the default, and verifying the binding between account and principal becomes the control — precisely the thing the design says is still open.
The paper’s open-questions list is a usable procurement test. Ask a registry, directory or agent-identity vendor four things: how control of the underlying domain or account is proven; whether records are authenticated at publication and how keys rotate and revoke; who is answerable for a wrong or disputed entry; and what a resolver is not asserting. If the answers are vague, what you have is a directory, not a trust layer.
Caveats
NandaIndex is a v0.1 design published for comment, with a reference implementation and no independent security review referenced in the material. The supporting quotes are endorsements of direction, not statements of integration, and the authors say so. The IETF-adjacent work in this area — including the federated-resolution draft from the same research group — remains individual submissions rather than adopted standards. Most importantly, the article is candid that identity, trust and global federation are not solved, and that a listing is not an authorisation; treat the design as a map of the problem, not as a control you can point an auditor at.
Sources
- R. Raskar, P. Chari, L. Muscariello, V. Pandey, S. Sharaf, J. Bu, K. Bharadwaj, “Making AI Agents Discoverable Without Owning a Domain”, MIT Media Lab, 24 October 2026 — media.mit.edu.
- A Global Switchboard for the Agentic Web (v0.1, request for comments), technical paper — nandaindex.org/paper.pdf.
- NandaIndex architecture, registration and reference implementation — nandaindex.org.
- R. Raskar et al., “Registry-Assisted Discovery for AI Agents and Workloads Without DNS Discovery Anchors”, `draft-raskar-agentic-web-federated-resolution`, IETF Internet-Draft — datatracker.ietf.org.
Note: the description above follows the published article and the v0.1 paper; no independent audit or third-party implementation of NandaIndex has been published.
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.