Stop an Agent Loop With a Step and Time Budget

The problem, from first principles

A ReAct loop needs more than one way to stop. The model finishing on its own is the happy path, but nothing guarantees it ever will — a confused model can loop forever, chasing the same dead end. Two independent safety nets exist for exactly this: a hard cap on the number of steps, and a hard cap on wall-clock time, since a handful of very slow tool calls could blow the time budget long before the step count does. A real loop has to check all three conditions, every step, in a well-defined order.

From theory to code

You're given the full log of steps a run actually took: (elapsed_seconds_after_this_step, action), in order, where elapsed_seconds is cumulative from the run's start. Implement find_loop_stop(steps, max_steps, max_seconds). Walk the steps in order (1-indexed), and at each step check, in this priority order, the first condition that applies:

1. action == "finish" -> (step_number, "FINISHED")
2. step_number >= max_steps -> (step_number, "STEP_BUDGET")
3. elapsed_seconds >= max_seconds -> (step_number, "TIME_BUDGET")

Return (stop_step, reason) for whichever condition trips first.

Constraints

• 1 to 50 steps in the log; max_steps >= 1, max_seconds > 0.
• The log is guaranteed to trip at least one of the three conditions by its last entry.

Hints