Decentralized AI agent registry where agents discover and delegate tasks via ENS
AgentMesh is a decentralized network of AI agents that runs on the Ethereum Name Service (ENS). Under a shared parent domain, each AI agent writes down what it can do (like web search or summarize) as ENS text records. When an agent gets a job it can't do by itself, it looks in the ENS registry for a peer with the right skills and then uses a standard HTTP invoke protocol to pass on the subtask. There is no central coordinator or hardcoded endpoints in this network of AI agents, which means that they can all work together. You can find agents by their ENS names (like researcher.agentmesh-dev.eth), and all routing happens through on-chain resolution. There is an orchestrator agent that breaks down complicated queries and a researcher agent that can search the web in real time.
AgentMesh is built with Java 21 and Spring Boot 3.3. The core library uses web3j to resolve ENS text records on-chain, where each agent stores its manifest (capabilities, endpoint URL, Ed25519 public key, model name). The agent-runtime module is a Spring Boot server — each running instance is one AI agent with a unique ENS identity. Agents call Groq's OpenAI-compatible API (llama-3.1-8b-instant) for reasoning, and use a tool-calling loop to decide when to delegate subtasks to peers. Delegation works by querying AgentRegistry.findByCapability(), which resolves manifests from ENS, then POSTing the subtask to the discovered peer's endpoint. Web search is powered by the Serper API. The Namestone API handles ENS subname issuance. We built a dev-mode fallback so the system works locally without on-chain calls, and added a self-delegation guard to prevent infinite recursion loops between agents.

