version: "1.0.0" name: multi-agent-a2a description: >- Help a learner compose agents as networked services using the open Agent-to-Agent (A2A) protocol — expose an agent as an A2A server and consume a remote agent as a peer. Use for Lesson 7 tasks: multi-agent orchestration and A2A client/server wiring.

Multi-agent & A2A

Guidance for composing agents over the Agent-to-Agent (A2A) protocol (Lesson 7).

Guardrails (always apply)

• Terminology: A2A (Agent-to-Agent) — an open protocol for agents to call each other as peers.
• Uses preview packages: agent-framework-a2a, a2a-sdk, and uvicorn (pinned in

requirements.txt with a preview note). Present these as preview, not GA.

• Config via .env: A2A_HOST, A2A_PORT, A2A_SERVER_URL. Models gpt-5.1 / gpt-5-codex.

Concepts a learner should grasp

1. Agents as services: instead of one process holding every agent, each agent can be exposed

over A2A and called remotely — enabling reuse, independent scaling and cross-team composition.

1. Server vs client: the server publishes an agent (with an agent card) over A2A; the client

discovers and calls it as a peer using A2AAgent.

1. Orchestration: a coordinating agent can delegate sub-tasks to specialised remote agents.

Worked example in this repo (validated live)

• lesson-7-multi-agent-a2a/a2a_server.py — exposes a Foundry agent over A2A (default port 9000).
• lesson-7-multi-agent-a2a/a2a_client.py — consumes the remote agent as a peer and prints its

response. This pair was validated end-to-end against a live Foundry agent.

How to help

• Start the server, then run the client; confirm A2A_SERVER_URL matches the server's host/port.
• Adapt the server's underlying agent (instructions/tools) rather than rewriting the A2A plumbing.
• Keep the preview caveat visible when advising on production use.

Validate

• python -m py_compile on both server and client.
• Client should reach the running server and return a real model response.

References

• lesson-7-multi-agent-a2a/README.md