Question

You are building an agent that lets Claude call a set of tools in a loop. In testing, you notice that for some ambiguous requests Claude keeps requesting tools repeatedly (stop_reason 'tool_use' every turn) and never produces a final text answer, causing runaway API costs. Which change to your application code best addresses this while keeping the tool-use loop functional for normal requests?

Accepted Answer

Set a maximum number of tool-call iterations in your loop, and after the limit is reached stop returning tool results and prompt Claude to answer with the information gathered so far.

Answer

Set tool_choice to 'none' on every request so Claude can never call a tool and always returns text immediately.

Answer

Remove the tool descriptions from the tool definitions so Claude is less likely to know when to invoke them.

Answer

Lower max_tokens to a very small value so each tool-use response is truncated before Claude can request another tool.