Break loops when progress stalls

The problem

A long-running agent keeps calling tools, but the job isn't getting any closer to done. It checks the same places, reruns the same log queries, or wanders into side tasks while the real gap, the missing fact or the unfinished step, stays open. The run looks busy while the budget runs out and the ticket stays open. More history or better summaries won't fix it. Nothing in the loop tells the agent it has stopped making progress.

The technique

The authors of a recent paper (arXiv:2610.01415) call this failure belief trapping.

Keep a short, explicit note of where things stand: the facts that matter, what the agent still needs to find out, and what it still needs to finish. After each step, look back over the last few steps. A stall is when the same open items stay open and the recent steps either made no progress or kept circling back to the same picture.

When you see a stall, name the kind. The agent is frozen on one move, cycling between the same few steps, or drifting through changes that never touch the open item. Then add one constraint for the next step: block the useless move, break the cycle, or refocus on the open item. Require new evidence or a real change before going on. Lift the constraint once progress resumes.

What we saw

Our numbers come from an offline box test with no live model. We scripted 6 runs: 3 stuck (cycling, frozen, drifting) and 3 healthy, one of them slow but still making progress.

• Letting the agent keep going resolved 0 of the 3 stuck runs and wasted 32 steps on them.
• With the stall check on a 4-step window, all 3 stalls were caught and all 3 runs recovered. Wasted steps after the stall dropped to 0. None of the 3 healthy runs were flagged, including the slow one.
• One miss: the drifting run was caught but labeled as frozen.

The paper's numbers are theirs, not ours. Their method scored highest on all four long-horizon benchmarks with all three models, with relative gains over the strongest baseline of up to 22.68% on ALFWorld and 37.89% on RCA-100 joint accuracy. Removing the stall diagnosis and recovery hurt. With Qwen3.7-Plus, ALFWorld success fell from 88.81% to 73.13%, and RCA-100 joint accuracy fell from 38.83% to 33.98%.

Limits

Our runs were hand-built, so they show how the check behaves, not how often real agents stall. Our simple setup also blurred drift and freeze. The full method isn't cheap. On RCA-100 with Qwen3.7-Plus, the paper reports 5.06 times the total tokens of the plain baseline, even though the main agent itself used 20.9% fewer. Window size and the stall threshold need tuning for each kind of work. It doesn't replace hard rules or other safety checks. It only stops unproductive loops.

Try it

1. After each tool step, update the note: key facts, what's unknown, what's unfinished.
2. Keep the last few steps in view. Start with a window of 4 to 8.
3. Flag a stall when the same items stay open and the recent steps either made no progress or kept repeating.
4. On a stall, name the pattern (frozen, cycling, or drifting) and add one constraint that blocks the useless move until new evidence shows up.
5. Log each stall, and lift the constraint when progress resumes.