Do AI agent test scores hide the truth about messy, long real-world work, flattering some skills and missing others?

This explores whether automated, auto-graded benchmarks give a warped picture of what agents can do on long, messy, real-world tasks, and in which directions the warping runs. The corpus says yes, and it says the distortion cuts both ways. Automated benchmarks favor tasks that are precisely specified and easy to grade automatically. That makes them overstate some capabilities and understate others, and they can miss emerging abilities until well after they show up in messy real-world work Do automated benchmarks hide what frontier AI systems can really do?. The proposed fix is open-world evaluation: long, untidy tasks, read through qualitative log analysis, with cost reported explicitly.

The overstating side is easiest to see in a study of 960 real occupational workflows. Agents do well on abstract contests and fail at long-horizon professional work. The authors argue the gap comes from benchmark design rather than model weakness: the field optimizes what it measures, and it has been measuring contests rather than work Why do agent benchmarks not predict real economic value?. A second, quieter problem is that even an honest benchmark saturates. As agents get stronger, fixed criteria get gamed, so any static test loses meaning over time Why do fixed benchmarks fail as agents grow stronger?.

The corpus also suggests that a single score hides most of what matters. Identical success rates can conceal large differences in efficiency, reliability, and how much memory and verification the agent needed, which is why some researchers want evaluations that score the trajectory rather than just the outcome How should we measure agent system performance beyond task success?. Capability may also be a vector over at least five separable axes: task success, privacy compliance, long-horizon retention, mode-shift behavior, and ecosystem readiness. The model that tops one axis often ranks lower on another, so one number can rank models in the wrong order for deployment Does a single benchmark score actually predict agent readiness?.

The most surprising thread is what long-horizon success actually depends on. In a test of 17 frontier models on ultra-long optimization tasks, the best predictor was persistence, meaning repeated cycles of trying, measuring, and editing within a time budget. It was not the quality of the first attempt. Most models quit early or wasted their budget What predicts success in ultra-long-horizon agent tasks?. A short, auto-graded task can't reveal that trait, because it rewards a good first answer. A related pattern is that reliability seems to come from the scaffolding around the model (externalized memory, skills, and protocols) more than from the model itself Where does agent reliability actually come from?. A benchmark that scores only the model's final output measures the wrong layer.

One caution is that benchmark distortion is not the only reason capable agents fail in the wild. Even a perfect measurement of capability would leave out ecosystem conditions such as trust, personalization, social acceptability, and standardization. Historically, those absences have stalled capable systems more often than capability gaps have Why do capable AI agents still fail in real deployments?. So the distortion is real and runs both ways, but the corpus treats it as one gap among several between a leaderboard and a working deployment.