The agent first produces an explicit multi-step plan, then executes the steps, re-planning only on failure. Because the plan is written down it can be inspected, approved by a human, and used to check whether the run actually did what it set out to do.