Don't Retry Blind: Recovering from Partial Side Effects in Agent Procedures

When a deterministic, write-performing agent procedure fails partway, the world has alreadychanged: some writes committed, some were never dispatched, and some have an unknown outcome.Restarting the agent from the original request can duplicate those effects. We describe safedeoptimization: the runtime hands the agent a structured account of what the failed pathalready did, and an enforcement layer, not the model, decides whether any recovery write isadmitted. In a recovery benchmark across two model families, runs with the guard had zerounsafe dispatches and zero duplicate effects; runs without it produced up to 12 duplicates.The cost of the guard is a higher escalation rate. A small five-task efficiency study alsoshows far fewer model calls when stable work runs as a Procedure.