Definition
Plain language
A way of sorting different kinds of agent mistakes into named categories.
As stated in the literature
An empirical classification of agent failure modes (e.g., Life-Harness's action-realization, contract violation, trajectory degeneration, reasoning error) used to guide harness design.
Why it matters: Naming failure modes lets researchers design targeted fixes instead of treating all agent errors as one undifferentiated blob.
For example, an agent might fail because it picked an action the environment doesn't support, broke a stated rule, looped forever, or simply reasoned wrong — each a distinct category.