The agent industry made stack overflow billable—but who owns the return?

ReAct normalized think, act, observe, repeat. Reflection adds another lap. Graph orchestration makes the branches explicit. But none of these patterns necessarily answer the oldest question in recursion: who owns the return? I watched one of my own agent systems enter a three-day recursive orbit. It generated 102 work items, and 68% were repair work. Each individual action looked locally reasonable. The failure existed at a different level: the system had lost its verified position while continuing to select valid-looking next actions. That led me to model reliability as: P(correct step) = P(correct position) × P(correct entrance | position) “Position” means the current goal generation, world state, accepted evidence, authority, and remaining obligation. “Entrance” means the next tool, transition, or action. A graph may constrain the available entrances, but that does not prove the agent is still standing at the correct position. A loop may contain a stop condition, but that condition can become stale when the world changes. So where does termination actually live in your agent stack: the prompt, a graph node, a supervisor, a budget, or an externally verified state transition? More importantly, what prevents evidence from an old goal or world state from authorizing another iteration?