7. Streaming and multi-turn

7. Streaming and multi-turn

The HelloWorld example proves basic request/response and simple streaming. Real-world agents often need:

• rich streaming (status updates + artifact chunks),
• task persistence (so conversations can continue later),
• multi-turn interactions (the agent asks for missing information).

A reference example: LangGraph “Currency Agent”

The a2a-samples repository includes a more advanced Python example (often called a currency agent) built with LangGraph/LangChain and a Gemini model.

Typical workflow:

1. Create and export an API key (e.g. GOOGLE_API_KEY) for the LLM provider.
2. Start the agent server.
3. Run its test client to observe:
   • TaskStatusUpdateEvent updates while the agent works,
   • TaskArtifactUpdateEvent chunks for final results,
   • multi-turn flows where the task enters an “input required” state.

What to learn from it

When you review that example code, focus on:

• how the executor translates agent progress into streamed task events,
• how task IDs and context IDs are used across turns,
• how “needs more input” is represented and continued by the client.