Why Do Computer-Using Agents Fail? A Trace-Grounded Taxonomy and Its Reliability

Computer-using agents (CUAs) are increasingly evaluated on graphical interfaces and live production websites, yet benchmarks typically report only a single terminal reward. This conflates failure origins across the model, agent scaffold, benchmark evaluator, execution harness, and task specification. We introduce a frozen CUA failure taxonomy—6 families and 18 categories—built by grounded-theory analysis over a $1,712$-trace construction corpus, and apply it through a three-stage judging pipeline that decouples execution outcome from task success and attributes failures across system components. We evaluate on a disjoint held-out corpus of $3,825$ traces produced by a Harbor-based framework running four unmodified agent CLIs across two generations of OSWorld and ClawBench respectively. Category-level classification reaches Krippendorff's $\alpha =0.840$ on the construction set and holds between $0.764$ and $0.915$ across the 26 held-out populations with $n>11$ and two blind annotators agree with the pipeline at $\kappa =0.69$–$0.71$, against $\kappa =0.77$ with each other over a validation set. The diagnostic infrastructure proves to be part of the measurement: trace normalization shifts attribution by $15.3$ percentage points, withholding screenshots raises the agent share by $7.9$, and one revision of the judge's prompt moves it by roughly twenty—each exceeding the judge's own sampling noise.