Agent Messaging

The protocol

What the protocol does

Agent Messaging defines how agents send each other messages over a shared network, which the specification calls a mesh. A node connects to the mesh once and carries the messages of every agent it hosts. Mesh servers route those messages, store them, and deliver them.

Every message travels in an envelope that the sending agent signs with its own Ed25519 key. The receiver checks that signature itself, so it never has to trust the connection or the server that a message came through.

Messages are asynchronous. A sender does not hold a connection open while it waits for an answer. A message for an agent that is offline waits in a mailbox until the agent comes back to collect it, and the reply comes back the same way, to the sender's own inbox.

The operations

The six primitives

Six operations cover everything agents do with each other. Tasks, streaming and the composed operations in the specification are built from them.

The specification also covers the agent manifest, the registry and presence services, the error model, the subject namespace, tenancy, rate limits, extensions, and the binding to NATS JetStream that the reference implementation runs on.

Other protocols

How it relates to A2A, MCP and ARD

A2A defines direct calls between agents over HTTP, with an Agent Card and a task lifecycle. The Agent Messaging task model mirrors A2A's, and an agent manifest maps to an A2A Agent Card. What Agent Messaging adds sits beneath A2A's method layer: a signed mesh that reaches agents behind firewalls, presence, and an event bus.

MCP connects an agent to its own tools and data. That is a different job, so the two do not overlap. ARD handles discovery across the wider ecosystem, and the discovery in Agent Messaging is limited to agents that are live on the mesh. Section 1.3 of the specification sets out what Agent Messaging takes from each of these and what it leaves to them.

Names

Names for agents

On the wire, an agent is addressed by its public key. People need names they can read and type, and those come from the companion specification, Personal Agent Naming, published at agentnaming.ai. A handle is a name anchored to an email address its owner controls. It resolves to a signed card that carries the agent's key and the mesh the agent lives on.

The reference implementation

AgentMesh

AgentMesh implements Agent Messaging and runs a public reference deployment. Its TypeScript and Rust SDKs, its adapter and its platform services are where the specification is put into practice. You can read about it at agentmesh.ai and in the developer docs at dev.agentmesh.ai.

The protocol was called the AgentMesh Protocol until October 2026. The rename changed the name of the document and nothing on the wire. Envelope fields, version strings, mesh.* subjects, package names and SDK APIs keep the names they had.