Basis researchers synthesize verified S3 compensators for agents
Researchers at Basis detail their work on synthesizing rollback systems, known as compensators, to make coding agents safer when interacting with stateful APIs like Amazon S3. The authors use an agentic loop combined with Dafny-based formal verification and "opportunity" feedback to generate compensators that are both robust and operationally useful. Evaluation across held-out tasks showed that verification combined with positive feedback significantly increased completed-and-restored tasks while eliminating observed errors and state mismatches in the tests. Open article →