Position Paper: Crash-Safe Effects for Agents in Operational Systems
Abstract
Agents have begun to operate computer systems. They migrate databases, roll out configurations, and remediate incidents, and the process that runs them can die mid-plan. When it does, neither obvious recovery is safe. Retry re-executes steps against a world its own earlier actions already changed. Restart begins with no record of which effects became true. Our position is that every side-effecting tool call must carry an effect contract made of three fields, an effect identifier, an idempotency key, and a declared reversibility class. The runtime writes the contract to a durable intent log before the call and reconciles that log on restart. We implement the log as EffectLog, 39 lines of Python over SQLite, and evaluate it on a thirteen-step database migration under process-kill injection. Across 320 measured runs, naive recovery leaves duplicate rows in up to 66% of migrations and re-executes a destructive step in up to 59%. Idempotency keys remove the corruption but still take 14.1 relaunches per migration. Reconciling from the log completes in 3.5 relaunches, at 2.8 ms of logging per effect. An agent that cannot show what it did must not be allowed to try again.