Once — AI Agent Execution Safety
Once is an open-source project focused on one narrow reliability problem: consequential AI-agent actions can succeed externally while appearing to fail locally, causing a retry to repeat the real-world side effect.
Maintainer: Jamie Oswald
Research identity: Once Research
Repository: stringsofthemind-oss/once
Website: onceexec.com
Technical paper: Reliable Execution of Consequential AI Agent Actions Under Retries and Ambiguous Outcomes
Design principle
A timeout does not prove an external action failed. Once therefore tracks a stable logical operation across retries, binds effect-bearing data to that identity, preserves durable state, and reconciles against provider truth where possible.
Claim boundary
Once does not claim universal exactly-once execution, atomicity between arbitrary external systems and its ledger, or automatic protection for every provider and framework. When the truth of an ambiguous operation cannot safely be established, the intended safety posture is to fail closed rather than guess.
Public development
The source, examples, hostile-retry labs, integration guides, research material, and issue history are developed publicly in the GitHub repository. The project is pre-1.0 and its supported paths are expected to be described narrowly and updated as reproducible evidence grows.
Project identity · Support matrix · Provider matrix · Research