Permission to Retry shows why a saved agent plan cannot authorize another effect

A durable agent may resume after a timeout without knowing whether the external operation succeeded. An old checkpoint says what the agent intended, but does not prove the payment, message or mutation was absent. The study isolates dispatch permission, effect evidence and permission to reconcile so a resumed worker cannot infer authority from an obsolete local state snapshot.

The study executes 336 prospectively frozen synthetic histories across LangGraph and Temporal. Per host, a retained-effect-ledger policy passes 42 of 42, while cached retry, recheck, and stable-key approaches pass 30, 36 and 31 respectively. The provider ledger is part of the research harness, not proof that every commercial API can expose equivalent effect receipts. Results are per-policy test histories and must not be quoted as broad model success rates.

Retain external effect identifiers, reconciliation status and the right to retry separately. On unresolved outcomes, hold the reservation and escalate instead of repeating the side effect. This complements earlier authorization-retry Cases but evaluates a distinct host-conformance experiment and a specific four-policy comparison. Primary evidence.

Additional evidence and trust boundary

A checkpoint contains the workflow's view of a pending action, but not necessarily proof that a payment, deletion or other side effect committed. In an ambiguous timeout, repeating with the same application intent can cause a second effect unless the provider recognizes a stable idempotency key. The provider ledger and authorized reconciliation decide whether to complete, deny or preserve an unresolved reservation. A reported action success without an external receipt is weaker than a verified effect. Supporting record.