CSE Roadmap

Common tier · about 4 hours · 11 parts

Part 3 of 11 · 2 min read

Budget alone can't tell a stalled agent from a slow one

A hard step limit stops runaway cost, but it has no way to say whether the agent was stuck the whole time or just thorough.

Think of a parking meter that simply expires after an hour. It stops you from parking forever for free. But it tells you nothing about whether you actually finished your errand or just sat there the whole time doing nothing.

In plain words: a budget is a hard limit on steps, tokens, money or time, a stall is a period with no measurable progress, a loop is the agent repeating the same action without effect, and drift is the agent working on something other than the goal -- a budget alone catches none of these by name, it only guarantees the run eventually stops, lumping a genuinely stuck agent together with one that was simply given an impossibly hard task.

A worked example. With a verifier but only a budget as the backstop (no stall detection), 35% of all tasks end by hitting the budget -- some of those are impossible tasks that could never finish, others are agents that genuinely stalled in a loop or drifted off-task, and the budget-stop outcome can't distinguish between them. Adding explicit stall detection on top of the same verifier raises correct outcomes from 0.75 to 0.89 when no tasks are impossible, and the gap widens dramatically as more tasks become genuinely impossible: at 40% impossible tasks, budget-alone scores 0.45 correct against 0.88 for budget-plus-stall-detection.

The key lines. From code/mini-goals/goals.py:

# from code/mini-goals/goals.py
else:
    u = r.random()
    if u < p.p_loop:
        mode = "loop"
    elif u < p.p_loop + p.p_drift:
        mode = "drift"
    elif r.random() < p.p_progress and done < ceiling:
        done += 1; progressed = True
idle = 0 if progressed else idle + 1

Read it as: idle increments on every step where progressed is false. That happens whether the agent is looping, drifting, or simply has not gotten lucky yet on a hard task. The step counter in a budget, by itself, is exactly this same idle-like count with no label attached. That is why a budget can only report "time ran out." It can never say "here specifically is what went wrong."

Budget, stall, loop, drift

A budget is a hard limit on steps, tokens, money or time. A stall is a period with no measurable progress. A loop is the agent repeating the same action or message without effect. Drift is the agent working on something other than the goal.

Example: the same "ran out of steps" ending covering both a looping agent and a legitimately-impossible task is exactly the blind spot a bare budget has -- it correctly stops the bleeding, but it tells the person reading the outcome nothing about what actually happened inside the run.