Agent-to-Agent (A2A) is a protocol for agents to talk to other agents across organizations and tech stacks. An agent card describes the agent (name, description, skills, endpoint, version). The executor sends chat context to the remote agent; that agent uses its own LLM and tools and returns an artifact (result + description). An event queue keeps connections alive for long tasks. So one travel agent can call an airline agent, a hotel agent, and a car-rental agent, each with its own model and tools, then combine the results.
A2A conversation flow
A2A: agents talking to agents
A2A pieces
- Agent card — Name, description, skills, endpoint, version.
- Executor — Sends chat context to the remote agent; remote agent uses its own LLM and tools.
- Artifact — Result + description returned when the subtask is done.
- Event queue — Keeps connection alive for long-running tasks.
Example: Full trip booking
User: "Book a full trip to Honolulu — flight, hotel, car." The travel agent (A2A client) delegates to airline, hotel, and car-rental agents. Each does its job and returns an artifact. The travel agent compiles a single response for the user.