Interesting, do you have any repo links or other sources on your experiment ? also regarding prior stale state looping, don't you think the agent could detect that by itself if given a sub-task to monitor for it?