Agent Discovery

Agent discovery is the process by which AI agents find, evaluate, and connect with other agents and services across networks. It is the foundational challenge of the Internet of Agents: before agents can collaborate, negotiate, or transact, they must first be able to locate each other and verify what they can do. Discovery is to the agentic web what DNS was to the early internet — an unglamorous infrastructure layer that makes everything else possible.

The problem is not trivial. Unlike web pages, which sit at static URLs, agents are dynamic. They spin up and shut down. Their capabilities change. They may operate across multiple registries controlled by different organizations with different trust models. A research agent looking for a specialized data-analysis agent needs to search across enterprise directories, public registries, and potentially decentralized networks — then verify that the agent it finds is actually capable of what it claims.

Discovery Architecture

Several architectural approaches have emerged. The NANDA Protocol from MIT Media Lab takes the most ambitious approach: a decentralized NANDA Index that functions as a universal agent directory. When an agent queries the Index, it routes the request to the relevant registry, validates cryptographic signatures via AgentFacts, and returns verified connection information. NANDA's Quilt architecture allows native, government, enterprise, and Web3 registries to interoperate — so an agent registered in a corporate Active Directory can be discovered alongside one registered on a blockchain-based registry.

Cisco's AGNTCY Directory takes a production-focused approach, implementing NANDA's DID-based schema in an enterprise-ready format. It demonstrates how academic research from MIT Media Lab can bridge to enterprise adoption through decentralized agent verification services.

At the protocol level, MCP handles discovery within its own ecosystem — an agent can discover available MCP servers and their tool capabilities dynamically. Google's A2A protocol provides discovery for agent-to-agent interactions. NANDA's contribution is the meta-layer: discovery across these protocols, so an MCP-native agent can find and connect with an A2A-native agent through protocol translation.

Semantic Matching

Name-based lookup is insufficient for agent discovery. An agent searching for "help with financial analysis" needs semantic matching — understanding that a "tax compliance agent" or a "portfolio risk assessor" might be relevant even if neither matches the search string. Advanced discovery systems use capability embeddings and behavioral history to match agents based on what they can actually do, not just what they call themselves. This connects to broader work in recommendation systems and AI search, but applied to agent-to-agent interaction rather than user-facing content.

Scale

The NANDA Index is currently hosted at 15 universities and partner institutions worldwide. The design target is trillions of agents — several orders of magnitude beyond today's deployments, but consistent with a future where every API, microservice, and autonomous function is wrapped in an agent interface. At that scale, discovery becomes a distributed systems problem as much as an AI problem, requiring the same kinds of caching, routing, and consistency guarantees that power the existing internet's infrastructure.