How idle LLM parent agents fall into costly polling loops — and how to fix it

When a parent AI agent delegates a task to a worker and has nothing else to do, it can fall into a repetitive polling loop, repeatedly waiting 60 seconds at a time just to learn whether the worker has finished. Each such wait triggers a full model activation, forcing the parent to re-read its entire conversation history, which in one measured run consumed over 10 million tokens across 89 waits. The root cause is structural: on platforms like Codex, a completed worker cannot directly wake an idle parent, so a 'do not poll' rule has nothing to enforce. Three potential fixes exist — extending the wait interval to reduce activation frequency, using a single long wait timed to the worker's expected duration, or deploying a zero-token watchdog script that monitors the worker silently and wakes the parent only once upon completion. The zero-token watchdog is highlighted as the most efficient solution, eliminating model token costs during the waiting period entirely.