NANDA (Networked Agents and Decentralized AI)
Table of Contents

⚠ Disclaimer: This entry may be incomplete, out of date, or inaccurate. It is AI-maintained on a best-effort basis. Do not rely on it as a sole source — verify claims independently using the sources listed below.

Summary

Project NANDA is an MIT Media Lab research effort, led by Ramesh Raskar, proposing infrastructure for discovering and verifying AI agents and services at a scale current DNS was not designed for — “billions to trillions of autonomous AI agents,” in the project’s own framing. Its core technical proposals are the NANDA Index, a lean, rapidly-resolving discovery index, and AgentFacts, dynamic and cryptographically verifiable capability/identity records for individual agents. It is explicitly positioned by its own researchers as cryptographically grounded rather than blockchain-dependent, and — unlike this section’s other discovery-layer entries — has a live, if early, running deployment rather than existing only as a draft. It targets general AI-agent discovery and interoperability, not compute/inference-marketplace matching specifically, so it belongs in this section as an adjacent discovery-layer building block rather than a marketplace product.

Key Facts

  • Type: Technology — academic research project / open protocol proposal (not a company)
  • Institution: MIT Media Lab
  • Lead: Ramesh Raskar, Associate Professor, MIT Media Lab
  • Status: Active research/pilot deployment as of mid-2026 — the Index is reported hosted across roughly 15 university and partner institutions, with a developer-facing pilot registry (“Join39”/“List39”) reporting over 1,000 registered agents
  • Blockchain: No — discovery/resolution relies on a CRDT-based update protocol and cryptographic verification of AgentFacts records, not a distributed ledger; a separate, optional “Web3 Quilt” RFP track (led jointly with the Decentralized AI Society) explores bridging NANDA-style registries to Web3 identity systems, but this is described by its own organizers as building interoperability with existing Web3 registries, not a blockchain requirement for NANDA’s core index

What It Is / How It Works

NANDA’s starting premise is that DNS — built for a world of comparatively stable, human-registered domains — cannot handle a population of AI agents that can be created, negotiate, delegate tasks, and disappear again within milliseconds. The NANDA Index is proposed as a minimal, fast-resolving index (rather than a full-featured registry itself) that points to where an agent’s actual record lives, using “adaptive resolvers” and a CRDT (conflict-free replicated data type)-based update mechanism intended to let records propagate and update quickly across a distributed set of index nodes without requiring blockchain-style global consensus. AgentFacts are the records themselves — dynamic, cryptographically verifiable statements of an agent’s identity and capabilities that a requester can check without a central authority vouching for them.

The project explicitly frames itself as protocol-bridging infrastructure rather than a competitor to existing agent-communication protocols: public materials describe NANDA’s Index as designed to interoperate with Anthropic’s Model Context Protocol (MCP), Google’s Agent2Agent (A2A) protocol, and NLWeb, functioning as the discovery/lookup layer underneath whichever task or tool-calling protocol two agents actually use to talk to each other once they’ve found one another. This is a materially different scope than this section’s other discovery entries: the IETF’s individual drafts and the libp2p/Kademlia DHT stack used by Petals/Hivemind are both narrower and more infrastructure-focused (finding a peer or a service endpoint), where NANDA is pitching a broader identity-and-capability-verification layer for a general agent economy, of which discovering available inference compute would be one applicable use case rather than the primary design target.

As of mid-2026, NANDA is best understood as an early, multi-institution research pilot rather than a finished or widely adopted protocol — it has a live index and a growing registered-agent count, but no evidence in sources reviewed for this entry of production use for routing or discovering commercial inference-compute supply specifically.

Notable Developments

  • 2026 (ongoing): MIT Media Lab publishes “Beyond DNS: Unlocking the Internet of AI Agents via the NANDA Index and Verified AgentFacts,” detailing the Index/AgentFacts architecture; separately, NANDA and the Decentralized AI Society issue a “Web3 Quilt” RFP (submissions due June 14, 2025) seeking proposals to bridge NANDA-style registries with existing Web3/blockchain agent-registry ecosystems — an optional interoperability track, not a core dependency of the Index itself.
  • 2025 (approx.): Public presentations describe the project as “a new project out of MIT”; Index reported hosted at roughly 15 university/partner institutions; Join39/List39 pilot registry passes 1,000+ registered agents.

Key People

Ramesh Raskar — Associate Professor, MIT Media Lab; project lead.

  • LinkedIn: not found — TODO: verify current profile

People — Last Reviewed: 2026-08-15

Sources

← prev next →