Running Agents

What it demonstrates

The base Agent gives you one verb — run() — with a few dials:

1. Non-streaming: result = await agent.run(prompt) → one AgentResponse
2. Streaming: agent.run(prompt, stream=True) → a ResponseStream you async-iterate
3. Sessions: agent.run(prompt, session=session) → conversation state across turns
4. Run options: agent.run(prompt, options={...}) → per-call temperature / max_tokens / etc.

Defaults are set once at construction via default_options=; per-run options= merge over them and win on conflict.

The code

# 2. Streaming — iterate the ResponseStream, then finalize.
stream = agent.run("Tell me one fun fact about Amsterdam's canals.", stream=True)
async for update in stream:
    if update.text:
        print(update.text, end="", flush=True)
final = await stream.get_final_response()  # aggregated result from the updates above

Non-streaming run() returns an AgentResponse whose .text is the aggregated answer and .messages is the full produced list (tool calls, reasoning, usage — not just final text).

The gotcha

Streaming is run(..., stream=True), not a separate run_stream() method. It returns a ResponseStream: iterate it for AgentResponseUpdate chunks (use update.text), then await stream.get_final_response() for the aggregated AgentResponse — that reuses the collected updates and does not re-run. Also, tools and instructions stay their own keyword arguments; they don’t go inside the options dict.

Azure / MAF mapping

The runner agent is built on FoundryChatClient (Azure AI Foundry) with AzureCliCredential, and default_options={"temperature": 0.7, "max_tokens": 500} baselines every run. Per-run options= override those Foundry generation settings for a single call.

Run it

uv run tutorial/02-agents/06_running_agents.py — needs Foundry creds. It worked if you see four labelled blocks: non-stream, stream, session recall, per-run options.

Next: Agent Pipeline