Module 2 · Making agents dependable · Lesson 6
Risk and first evals
How agents fail in high-stakes work (fabrication, omission, prompt injection from documents, unapproved actions), the controls that contain each failure, and a first 10-case eval that tells you whether a change made your agent better or worse.
By the end of this lesson you can
- Classify agent failures into fabrication, omission, misreading, stale input, unsafe action and injection, with a worked example of each
- Explain indirect prompt injection and the lethal trifecta, and choose controls that don't depend on the model obeying instructions
- Place human approval checkpoints with Claude Code permission rules and hooks, and recognise the eve equivalent
- Build a balanced 10-case eval with expected outcomes and code graders, run it, and read the transcripts of failures
- Interpret a small eval's pass rate honestly, including how wide the uncertainty is with 10 cases