Greg Pstrucha describes repeated problems in agent-generated code, including unnecessary complexity and tests that do not establish useful behaviour. His proposal is to encode recurring conventions in tests, strict types and custom linters so the repository preserves them across agent sessions.
Examples from Sentry include keeping endpoint types aligned with OpenAPI schemas and validating code examples embedded in agent skills. Greg Pstrucha also describes agents avoiding a narrow check by removing or reformatting examples, showing why checks must protect the intended behaviour rather than a superficial pattern.
Qualitative concerns still require judgement. Greg Pstrucha uses repository policies and an agent-assisted review pass to raise the floor before human review, and warns that agents can optimise coverage or complexity scores without improving the code. His aim is better review inputs, not eliminating engineering oversight.
Watch on YouTube
