A monorepo comparing three Python agent frameworks, each implementing the same demo agent exposed via the A2A (Agent-to-Agent) protocol.

Demo task: Given a partial French children's rhyme ("3 petits chats / qui mangeaient des rats"), return the next line.

a2a-xp/
├── agents/
│   ├── ag2/          ← AG2 agent + docker-compose.yml (standalone)
│   ├── crewai/       ← CrewAI agent + docker-compose.yml (standalone)
│   └── langgraph/    ← LangGraph agent + docker-compose.yml (needs infra)
├── infra/            ← postgres + redis (for LangGraph)
├── client/           ← A2A test client
├── docker-compose.yml  ← root: composes everything with profiles
└── .github/
    └── copilot-instructions.md

• Docker + Docker Compose v2.20+
• OpenAI API key

cp .env.example .env
# Edit .env and fill in OPENAI_API_KEY

Each sub-project is independently runnable:

# AG2 agent only
cd agents/ag2
docker compose up

# CrewAI agent only
cd agents/crewai
docker compose up

# LangGraph agent (needs infra running first)
cd /path/to/repo
docker compose -f infra/docker-compose.yml -f agents/langgraph/docker-compose.yml up

# All agents + infra
docker compose up

# Specific agent only (using profiles)
docker compose --profile ag2 up
docker compose --profile crewai up
docker compose --profile langgraph up

# With all agents running:
docker compose run client

Uses uv for Python dependency management:

cd agents/ag2          # (or crewai, langgraph, client)
uv venv && source .venv/bin/activate
uv pip install -r requirements.txt
uv run uvicorn app.main:app --reload --port 8000

Once running, each agent exposes:

• GET /.well-known/agent.json — AgentCard discovery
• POST / — A2A JSON-RPC endpoint (message/send)

Each sub-project has a tests/ directory. Run with uv run pytest from the sub-project directory (no Docker, no real API key required — LLM is mocked):

cd agents/ag2    && uv run pytest
cd agents/crewai && uv run pytest
cd agents/langgraph && uv run pytest
cd client        && uv run pytest

Key findings:

• AG2 and CrewAI share the same a2a-sdk plumbing pattern — almost identical server bootstrap code. Difference is purely in the agent/crew definition.
• CrewAI's kickoff() is synchronous and must be dispatched via asyncio.get_running_loop().run_in_executor() to avoid blocking uvicorn.
• LangGraph has the richest A2A feature set out-of-the-box (streaming, push notifications, task history) but requires postgres + redis, making standalone development more complex.
• All three expose /.well-known/agent.json automatically via their respective A2A plumbing — no hand-rolling of the AgentCard HTTP endpoint needed.
• The a2a-sdk version 0.3.x uses message/send (not the older tasks/send); always verify the method name against the installed SDK version.