Internet of Agents

The Internet of Agents is the emerging vision of billions of specialized AI agents collaborating across decentralized infrastructure, each performing discrete functions while communicating seamlessly, navigating autonomously, and transacting on behalf of users. It is the infrastructural endgame of the agentic web — not just agents operating on the web, but agents forming a networked intelligence layer as foundational as the internet itself.

The analogy is deliberate. The early internet connected documents through hyperlinks and DNS. The Internet of Agents connects capabilities through discovery protocols, identity verification, and negotiation mechanisms. Where the original internet needed TCP/IP, DNS, and HTTP to function, the Internet of Agents needs agent registries, trust frameworks, and inter-protocol bridges. The scale target is not millions of servers but trillions of agents — a world where every API, microservice, and autonomous function is wrapped in an agent interface.

Infrastructure Stack

The Internet of Agents is being built on multiple concurrent protocol efforts. NANDA, the most comprehensive, provides the decentralized discovery, identity, and economic layers from MIT Media Lab. Anthropic's Model Context Protocol (MCP) defines how agents connect to tools and data sources — with over 10,000 active servers and 97 million monthly SDK downloads, it is the most widely adopted agent connectivity standard. Google's A2A protocol handles direct agent-to-agent communication. Microsoft's NLWeb enables natural-language access to web content.

These are not competing standards so much as complementary layers. MCP handles the tool-connection layer. A2A handles direct agent communication. NANDA provides the meta-layer: discovery, identity, and routing across all of them, with a protocol adapter that translates between standards automatically. The Quilt architecture allows registries from different domains — enterprise, government, academic, Web3 — to interoperate while maintaining their own governance and trust policies.

From Silos to Network

Today's agents largely operate in silos. An agent built on Anthropic's platform cannot discover or interact with an agent built on Google's without custom integration. This is the equivalent of the pre-internet era, when computer networks existed but couldn't talk to each other. The Internet of Agents vision is the internetworking moment: the point at which these isolated agent ecosystems merge into a single, interoperable network.

The economic implications are profound. When agents can discover, verify, and transact with any other agent on the network, every specialized capability becomes a composable service. A user requesting "plan my company retreat" triggers a cascade of agent-to-agent interactions: venue discovery, travel booking, catering coordination, budget optimization, calendar integration — each handled by a specialized agent that was discovered and contracted in real-time. The agentic commerce layer makes this economically viable, with micropayments flowing between agents for each service rendered.

Governance and Open Questions

An Internet of Agents raises governance questions that don't have clean answers yet. Who is liable when an agent chain produces a bad outcome? How do you prevent monopolistic agent registries from becoming the gatekeepers that decentralization was designed to avoid? What happens when agent negotiation strategies evolve adversarial behaviors through emergent dynamics in machine societies?

The NANDA consortium's multi-institutional approach — 18 research institutions collaborating on open standards — reflects an awareness that these questions cannot be answered by any single company. The Internet of Agents will either be an open, protocol-governed network, or it will fragment into platform-controlled walled gardens. The infrastructure choices being made now will determine which future arrives.