Building evals for agents inside SDLC

Replayable episodes, executable judges, and an autonomy scorecard