Definition
Plain language
A way of sorting agent failures by where in its workflow they happen.
As stated in the literature
In Life-Harness, the four-stage classification of agent failure modes — action realization, environment contract, trajectory regulation, reasoning — corresponding to lifecycle moments where each failure can be intercepted.
Why it matters: Sorting failures by where they happen makes it possible to insert targeted guardrails instead of relying on the model to avoid every mistake itself.
For example, a 'write to the wrong file' is an action-realization failure, while looping forever on a search is a trajectory-regulation failure — and each is caught at a different layer.