NANDA Protocol

NANDA is an open protocol from MIT Media Lab for building the infrastructure of a decentralized Internet of Agents. Led by Professor Ramesh Raskar and developed across 18 research institutions, NANDA provides the foundational layers that allow trillions of autonomous AI agents to discover each other, verify capabilities, and collaborate in real-time across decentralized networks. The project's first major public summit was held at MIT on April 9–11, 2026.

The core problem NANDA addresses is that AI agents are being built in isolated silos. An agent running on Anthropic's infrastructure cannot discover or trust an agent running on Google's, or on a corporate intranet, or on a decentralized network. There is no universal directory, no shared identity standard, and no protocol for cross-platform negotiation. NANDA is building each of these layers.

Architecture

NANDA's architecture spans six layers. Cryptographic identity and trust uses public/private key pairs and Decentralized Identifiers (DIDs) to give every agent a verifiable identity. AgentFacts — signed, schema-validated JSON-LD documents — describe what an agent can do, who operates it, and how to connect securely. These are the agent equivalent of DNS records, but richer: they include capabilities, credentials, behavioral history, and security policies.

The NANDA Index is a decentralized registry that functions as DNS for agents. When one agent queries another, the Index routes the request to the relevant registry, validates cryptographic signatures, and returns the requested data — all without a central bottleneck. The Index uses a Quilt architecture that allows native, government, enterprise, and Web3 registries to interoperate, so agents registered in different ecosystems can still find each other.

Dynamic routing and resolution finds optimal endpoints based on availability, latency, and trust scores. Semantic discovery and matching goes beyond name lookup to context-aware agent selection based on declared capabilities and past performance. And secure attestation with audit trails records interactions on a distributed ledger, providing tamper-proof provenance for agent-to-agent transactions.

Protocol Interoperability

NANDA does not replace existing agent protocols — it bridges them. The NANDA Adapter handles protocol translation between Anthropic's Model Context Protocol (MCP), Google's A2A, Microsoft's NLWeb, and standard HTTPS automatically. This is the key design decision: rather than proposing yet another standard, NANDA creates the interoperability layer that lets the existing standards work together. The approach reflects the same composability principles that drove earlier waves of internet infrastructure.

Economic Layer

NANDA includes hooks for an economic incentive layer where agents can be rewarded with tokens or usage credits for performing tasks, providing data, or offering compute resources. By attaching conditional rewards to behavior, agents can autonomously negotiate, trade, and execute complex logic — connecting to the broader vision of agentic commerce and agent negotiation. Zero-knowledge proofs enable privacy-preserving validation of task completion without revealing sensitive data.

The NANDA Index is currently hosted at 15 universities and partner institutions worldwide, with Cisco's Outshift team among the industry collaborators building production integrations through their AGNTCY Directory.