695 real agent failures traced to how much freedom agents were given

Researchers catalogued 695 real incidents where AI agents failed, then built a 12-category taxonomy from the bottom up. The central finding: how much freedom an agent was given at deployment shaped the kind of failure that followed. Model capability was not the deciding factor. The taxonomy covers failures across the full arc of agent work: before the agent starts, while it runs, and after it finishes. Failures before the start include missing or wrong information handed to the agent at launch. Failures mid-run include the agent acting beyond its brief or failing to surface that it is stuck. Failures at the end include work handed back without enough for a human to judge it. For product teams, the practical implication is direct. Decisions made during setup, about reach, what the agent is allowed to do, and when it should stop and ask, do more to prevent failure than model selection. Teams designing agent products should treat those setup decisions as the primary design surface.