This is a sample field note. Agents fail in the gaps between tool calls. Eval the trajectory, not only the final sentence.
Build a scenario pack early
- Include happy paths, low-confidence cases, and hostile inputs.
- Assert required tools were called with permitted arguments.
- Score escalation behaviour when confidence is below threshold.
Gate every prompt or model change
Run the pack in CI before promotion. Track regressions by scenario class so teams see which failure mode they just introduced — not only a single accuracy number.
Keep humans in the audit path
Store tool calls, approvals, and overrides with enough context to reconstruct a bad day. Without that trail, production incidents become folklore.





