Agent loops carry three critical risks: they can enter infinite loops (a programming hazard), drift as agents coordinate and go off-course, and accumulate token costs as the loop continues executing.