AI agents fail in a predictable set of ways once they can act, and predictable means checkable before promotion. Get the working version of the catalogue and score your agent before its next promotion review.

​

The catalogue. All 18 failure modes, grouped by the four causes behind them, with how each shows up in a run, a detection test you can run, and the controls that catch it.

The one-page scoring sheet. Score the control that actually runs for each mode, from 0 (not considered) to 4 (gated), for one action at its target autonomy level. Critical gaps stand out at a glance.

The scoring workbook. The same sheet in Excel: set the target level and it applies the promotion bar and gives the verdict, plus every source behind every row.