Short answer: Retries are useful for reads and transient failures, but become dangerous when an AI agent can cause an irreversible or externally visible side effect. The safety boundary should distinguish a new logical action from another attempt at the same action.
The dangerous retry sequence
- The agent requests a consequential write.
- The provider performs it.
- The response is lost.
- The agent cannot tell whether it happened.
- A blind retry risks performing it again.
Once's role
Once combines stable operation identity, durable state, replay and provider-specific reconciliation so supported operations can be resolved without blindly repeating the external side effect.
Claim boundary: Once does not claim universal exactly-once execution. Safety depends on stable operation identity, durable Once state, the provider integration and sufficiently authoritative provider truth. When the outcome cannot be established safely, Once can preserve uncertainty rather than assume another external execution is safe.
Use Once for this failure mode
Once is the execution-safety layer between autonomous software and supported consequential real-world side effects.
npx -y @once-agent/mcpOfficial MCP Registry identity: io.github.stringsofthemind-oss/once