Responses API infinite loop of tool calls

[warped marsh]

The task is done, but model requests tool calls non stop.

[fluid charm]

Hi @warped marsh. I'm kapa.ai, your friendly support bot. I will try to answer your question by searching through the knowledge base. :wave:

---

This behavior is usually caused by how tool use is configured, not by the Responses API itself.

From the official docs and examples, there are three main ways to stop infinite tool-call loops:

1. Limit how many tools can be called per response

   The Responses API supports max_tool_calls, which hard-stops further built‑in tool calls once the limit is reached. Any additional attempts by the model to call tools are ignored, which prevents runaway loops (for example with image_gen). [Image tool answer]

2. Constrain tool‑use behavior in your agent / orchestration

   In the openai-agents-python framework, tool behavior is explicitly configurable:

   • Default: "run_llm_again" – run tools, then call the LLM again with results.
   • "stop_on_first_tool" – treat the first tool’s output as the final answer.
   • StopAtTools(stop_at_tool_names=[...]) – stop once a specific tool has been called and use its output as final. [Tool use behavior]

   These options are designed specifically to avoid the pattern “tool → LLM → tool → LLM …” when the task is already complete.

3. Reset tool choice to avoid forced re‑calling

   The agents framework automatically resets tool_choice to "auto" after a tool call via Agent.reset_tool_choice (default True) to prevent infinite loops of tool usage. If tool_choice is pinned (e.g., always a specific tool), the LLM can be forced to call tools again and again. [reset_tool_choice]

“To prevent infinite loops, the framework automatically resets tool_choice to "auto" after a tool call… The infinite loop is because tool results are sent to the LLM, which then generates another tool call because of tool_choice, ad infinitum.” [Tool use behavior]

---

What you can do concretely

• Set a hard cap with max_tool_calls in your Responses request. [Image tool answer]
• If you’re using openai-agents-python, ensure:
  • reset_tool_choice is left as True, or explicitly set to True. [reset_tool_choice]
  • tool_use_behavior is "stop_on_first_tool" or StopAtTools(...) when the tool output itself should be the final answer. [Tool use behavior]
• In custom loops (your own orchestration), add:
  • A max iterations guard (like the 5‑turn loop in the Node.js agent example). [Node agent loop]
  • A check that breaks once you have a satisfactory tool result instead of always sending it back to the model.

If you share your current request body or orchestration pattern, I can map these controls directly onto it using the same patterns from the docs.

Note: you can ask me a follow up question by @ mentioning me again :speech_balloon: