Agent2Agent Protocol (A2A)

Agent2Agent (A2A) is an open protocol from Google for letting AI agents built in different frameworks and languages talk to each other. Announced April 2025, with production-ready support arriving in adk-go 1.0 in March 2026.

The problem

If you build one agent in Go and another in Python, they can't naturally cooperate. Each framework has its own internal message format, execution model, and state management. Without a shared protocol, "multi-agent system" really means "multiple agents in the same framework." A2A gives agents a common language regardless of what they're built with.

How it works

A2A defines two roles: a client agent that formulates tasks, and a remote agent that executes them. Communication happens over standard web infrastructure — HTTP, Server-Sent Events (SSE), and JSON-RPC.

Four core capabilities:

Capability discovery. Agents publish "Agent Cards" — JSON documents describing what they can do, what inputs they accept, what they produce. A client agent reads the card to figure out whether a remote agent is useful for a given task.

Task management. Tasks are first-class objects with defined lifecycles. A task can complete immediately or run for hours/days (important for workflows involving human-in-the-loop approval). The client gets real-time status updates via SSE.

Message exchange. Agents send messages back and forth with structured "parts" — text, data, artifacts. Context flows with the messages so the remote agent understands what it's working on.

Content negotiation. The parts system lets agents negotiate format. One agent might produce an iframe, another prefers plain JSON. They work it out through the protocol rather than requiring all agents to speak the same output format.

Relationship to MCP

A2A complements Anthropic's Model Context Protocol (MCP) rather than competing with it. MCP gives agents access to tools and context — it's about what an agent can do. A2A is about agents coordinating with each other — it's about how agents work together. You might use MCP to connect an agent to a database, and A2A to have that agent collaborate with a scheduling agent built in a different language.

Adoption

Over 50 technology partners back A2A, including Atlassian, Salesforce, SAP, LangChain, MongoDB, and PayPal. Major consulting firms (Accenture, Deloitte, McKinsey) are also involved, which signals enterprise adoption intent.

The spec is open source on GitHub.