The following is based on an internal talk given at Vercel. We're sharing it publicly because the problem it describes isn't unique to us, and the framework is useful for any team shipping with agents. Coding agents generate code at unprecedented speeds. In the hands of disciplined engineers, they are a productivity multiplier. But without rigorous judgment, they are a highly efficient way to ship bad assumptions directly to production. When teams deploy agent-generated code blindly, the fallout can be immediate and severe. A flawless-looking pull request can ship a query that passes tests, but scans every row in production.

Retry logic that seems correct can cause a thundering herd on a downstream service. And a cache with no TTL can quietly grow until Redis dies. Green CI is no longer proof of safety. In an agentic world, passing CI is merely a reflection of the agent's ability to persuade your pipeline that a change is safe, even if it will immediately degrade your infrastructure at scale. Agent-generated code is dangerously convincing. It comes with a polished PR description, passes static analysis, follows repository conventions, and includes reasonable test coverage. On the…