Three-Gate Framework Proposed to Catch Failures in Agent-Written Code Patches
A software engineer has outlined a three-gate validation system designed to catch common failure modes introduced by AI agent-generated code patches before they are merged. The first gate uses property-based testing to verify structural invariants — such as ensuring a rewritten function never returns negative values — rather than relying solely on example-based tests. The second gate pins test fixtures to a manifest file with checksums, so any silent data changes made by an agent trigger a merge-blocking alert requiring human approval. The third gate addresses flaky tests by quarantining them with a mandatory expiry date and a linked ticket, forcing developers to either fix, replace, or explicitly delete the test. The author argues that the ordering of these gates is critical, as each is designed to address a distinct failure mode: wrong logic, altered test data, and nondeterminism respectively.
