How does an agent know when to stop?

Layers of stopping

1. Natural termination: the model returns a response with no tool call, or calls a dedicated final_answer / submit tool. This is the cleanest pattern because the final output is structured.
2. Hard limits (always in code, never just in the prompt):
   – max_iterations (e.g. 15–25 for most tasks)
   – token / cost budget per run
   – wall-clock timeout
3. Loop detection: the same tool with the same arguments N times, or no new information over K steps → break out or inject a hint.
4. Success verification: for some tasks the environment can confirm success, e.g. tests pass, the record exists in the DB. Prefer this over the model's own claim that it's done.
5. Graceful failure: when a limit is hit, return a useful partial result and explanation, not an exception.

Prompt-side support: tell the agent what "done" means and that it's acceptable to stop and say "I couldn't find X; here's what I tried." Many agents loop because they think giving up is forbidden.

Common mistakes

• Relying only on the model to stop.
• Raising an error at max iterations and losing all the work done so far.

Follow-ups to expect

• Scenario: the agent keeps calling the same tool. (See Q18.)

Related

• Previous: Q11. What is human-in-the-loop (HITL), and where should you place it in an agent?
• Next: Q13. What is agentic RAG, and how does it differ from classic RAG?
• LangGraph
• LangChain