Agent runtime

A2A (Agent-to-Agent Protocol)

A protocol for agents to discover, describe, and call each other directly — Google's open spec for letting independent agent systems collaborate without each one becoming a tool for the other.

Operating principle

Production AI is not a prompt. It is a system of context, tools, permissions, traces, evals, and feedback loops.

What it is

A2A — short for Agent-to-Agent — is a protocol for one AI agent to find and work with another agent it doesn't already know. Google introduced it in 2025 as an open standard. The idea: an agent advertises what it can do through an Agent Card; another agent reads the card, decides whether the first agent is useful, and sends it a task. The two agents can then exchange messages, share state, and converge on a result — without one being hardcoded inside the other.

Why it matters

Most agent systems today are monoliths or closed ecosystems. A2A is the early attempt to let agents built by different teams (or different companies) work together at runtime — closer to how humans cooperate across organizations than to how APIs integrate today. Whether the standard catches on is open, but the problem it addresses is real.

How it works

An A2A server publishes an Agent Card (a structured description of its capabilities, schemas, and trust requirements). A client agent discovers cards, picks one, and sends a task; the server streams back updates. Authentication and trust handshakes happen at the protocol layer. Related to the Model Context Protocol (MCP) but different in scope: MCP exposes tools to a single agent; A2A connects agents to each other.

Related resources