Execution Budgets: Steps, Time, Tokens, Money

Forge's harness stops a run at 30 steps. That limit sounds complete, so test it with some arithmetic. The sandbox lets each command run for up to 5 minutes, so 30 steps permit 150 minutes of commands alone.

The step count does not limit tokens either. Under this course's standard inputs, each step adds a 1,500-token tool result to the history. Suppose instead that each result is a 10,000-token test log. The run's total then passes one million tokens at step 14, which is 16 steps before the step limit.

A step limit counts one thing, and a run can fail in ways that use few steps