Multi-step tools — when one tool isn't enough (step 9/9) · agent loops

Checkpoint

One last thing before we move on. pass this to mark the lesson done, or skip and keep moving. hop to the next when you're ready.

Final drill. Build a research agent with three tools and parallel tool calls. Write run_research_agent(question, max_iterations) that:

Maintains a messages list, starting with one user turn.
Loops up to max_iterations times, calling fake_model(messages).
On end_turn: return {"ok": True, "answer": <text>, "rounds": <iterations used>, "tools_called": <total tool runs across all turns>}.
On tool_use: handle ONE OR MORE tool_use blocks per turn. For each, dispatch through TOOLS. If the tool name isn't in the registry, the tool result content should be "unknown tool: <name>" (don't crash). Each tool you successfully run counts toward tools_called. Append the assistant turn + one user turn with all tool_result blocks.
If max_iterations is exhausted without end_turn: return {"ok": False, "answer": "capped", "rounds": max_iterations, "tools_called": <count so far>}.

Two cases run for you. Expected output:

ok=True rounds=3 tools=3 answer=found 2 sources on ramen
ok=False rounds=2 tools=2 answer=capped

Checkpoint

One last thing before we move on. pass this to mark the lesson done, or skip and keep moving. hop to the next when you're ready.

Final drill. Build a research agent with three tools and parallel tool calls. Write run_research_agent(question, max_iterations) that:

Maintains a messages list, starting with one user turn.
Loops up to max_iterations times, calling fake_model(messages).
On end_turn: return {"ok": True, "answer": <text>, "rounds": <iterations used>, "tools_called": <total tool runs across all turns>}.
On tool_use: handle ONE OR MORE tool_use blocks per turn. For each, dispatch through TOOLS. If the tool name isn't in the registry, the tool result content should be "unknown tool: <name>" (don't crash). Each tool you successfully run counts toward tools_called. Append the assistant turn + one user turn with all tool_result blocks.
If max_iterations is exhausted without end_turn: return {"ok": False, "answer": "capped", "rounds": max_iterations, "tools_called": <count so far>}.

Two cases run for you. Expected output:

ok=True rounds=3 tools=3 answer=found 2 sources on ramen
ok=False rounds=2 tools=2 answer=capped

full-screen editor opens — close anytime to keep reading.

Multi-step tools — when one tool isn't enough — step 9 of 9