Song Xu tests what happens when an agent changes a destination but loses confirmation. An agent is a model running multi-step tasks with tools. Preventing a repeated action and establishing whether it finished require different safeguards.
The runtime records intent before dispatch, gives actions stable identifiers and asks the destination to suppress duplicates. After interruption, it checks destination evidence to resolve the outcome. A separate destination database records what actually happened.
- On each of Windows and Linux, 7,513 fault/control trials reuse 96 captured model responses from 24 constructed cases. They are not independent live-model decisions.
- In 200 known crash-and-replay comparisons per platform, the full setup repeats zero effects. Each simpler baseline repeats effects in 117 comparisons.
- Disabling only destination checks preserves duplicate prevention but leaves all 100 known post-effect crashes ambiguous. A later-designed response-loss supplement finds the same distinction, using the same inputs.
- Removing a bundle of ledger capabilities leaves 170/170 effects without terminal ledger records. That ablation cannot isolate one component’s contribution.
Durability has a cost. The full setup runs at roughly one-seventh of the permissive baseline’s measured operation rate. Evidence-write faults leave 75/100 trials unresolved on each platform because the runtime refuses to finalize without evidence.
Caveat: This is one team’s controlled SQLite implementation; code, raw records and aggregate analysis tables are not publicly deposited. The two platforms demonstrate consistency, not independent replication.
Record intent, make retries safe, then verify the exact destination. A durable action ID cannot replace that last check.