Agent loops need runtime controls: budgets, deadlines, cancellation, and idempotency

The article argues that production failures in AI agents usually stem from the execution loop rather than model quality, such as runs that spin for long periods, duplicated side effects, or cancellations that fail to stop workers. It recommends treating the loop as a runtime with admission control, deadlines, cooperative cancellation, and at-most-once side effects. Agent runs should have separate limits on steps, wall-clock time, tool calls, and cumulative tokens or cost, since per-step costs rise as conversation history grows. Cancellation must also reach inside tool calls through propagated deadlines and explicit timeouts on every blocking operation. Finally, retried writes should be prevented with deterministic idempotency keys stored alongside the effect, and third-party APIs without idempotency support should use an outbox-style record-and-reconcile approach.