The five steps
- One tool call 8 min
- Keep going until it answers 9 min
- When a tool fails 10 min
- A step budget 10 min
- Several calls in one turn 11 min
Write the loop at the center of every tool-using agent: send the conversation to a model, run the tools it asks for, feed back the results and the errors, stop it gracefully when it runs out of steps, and handle several calls in one turn without running the same one twice.