Agentic tool use with Claude XML

Combine XML scaffolding with the API's tool-use feature for reliable agent loops.

How to apply the pattern

1. Use real tools via the API, not XML-defined ones. tools=[...] in the API call is stronger than asking Claude to output <tool_call> tags.
2. Wrap tool observations in <observation> tags. When you re-inject tool results, give them clear boundaries. The model treats them as evidence, not instructions.
3. Have Claude plan in <plan> before acting. On the first turn, ask for a numbered plan. Subsequent turns can re-plan with <updated_plan>.
4. Use <final_answer> as the stop signal. Your agent loop terminates when this tag appears.

Worked example

<instructions>
You are an agent that helps users plan trips. You have tools available.

On each turn:
1. Reason in <thinking> tags about what to do next.
2. Either call a tool, or produce a <final_answer>.
3. When you call a tool, the result will appear in <observation> tags on
   the next turn.

Do not include <observation> tags yourself — those come from the system.
</instructions>

<task>Find me a flight from SFO to NYC next Tuesday under $400.</task>

Tips

• Keep tool descriptions short and example-driven — Claude calls tools more reliably when their JSON schema includes an example argument.
• Bound the agent loop. Hard cap on turns prevents runaway costs.