Don’t let the coding agent grade its own work
When a coding agent finishes a change, it is tempting to ask that same agent to run the final checks and report whether the work passed.
Kogen deliberately separates those responsibilities.
The Developer can still run focused tests while implementing a feature. That is how it reproduces failures, checks repairs, and investigates unexpected results. But it cannot run the final verification that determines whether its work can proceed to Review.
Those checks belong to Kogen’s automated machinery. Their results are recorded against the exact code and Developer session being assessed. Review then belongs to a fresh agent.
We enforce this separation at runtime for the commands Kogen uses as final verification. If the Developer tries to run one, it is blocked before execution and directed toward focused tests instead. This is a bounded ownership mechanism, not a general-purpose shell sandbox.
The principle is simple: let a coding agent use tests to do its job, but do not let its own success report become the evidence that the job is finished.
← All notes