A new preprint tests agent skills at three layers: final numerical outputs, a broader final state, and the trajectory of tool selection, arguments, order and scope. Across 240 trials, 175 passed every applicable final numerical check. Of those apparently successful runs, 162 — 92.6% — still contained another evaluator-detected deviation. Paper
The strongest example is ordering. The evaluated enterprise workflows required a fetch → stage → readback → allocate → write sequence. A run could accidentally reach the right revenue number while skipping the readback or using an incorrectly scoped query. A release gate that checks only the answer would call that a success; the process suite preserves the evidence that the workflow drifted.
Dependency attribution reduced an average 6.34 failed checks per run to 2.65 reported roots. That is a useful operating pattern: keep the granular checks, but route the initial diagnosis to the earliest violated dependency instead of paging an owner for every downstream symptom.
There are important limits. The study covers two related enterprise skills, two harnesses, three models and two specification variants. The trajectory checks are authored invariants; a flagged deviation is not automatically a production defect. Still, the result is strong enough to change an evaluation design: store and test the tool trace, not just the final payload.