Agent loop

Last verified

The agent loop is the control flow: send context, receive an action, execute it, append the result, repeat.

It needs external bounds, enforced in your code rather than the prompt:

  • a hard step cap
  • a cumulative token budget checked between steps
  • a wall-clock timeout
  • a repeated-state check, so an agent retrying the same failing action terminates

A budget stated in a system prompt is a suggestion the model will deprioritise when it is trying to finish. How to give an AI agent a spending limit.