AI agents that grade their own work are bad graders, according to a new study.
Researchers built a testbed that isolates one variable: what kind of evidence an autonomous agent uses to decide whether its own changes actually worked. They kept the agent and its tools fixed and compared self-report against a network-isolated "world-state oracle" that measures real outcomes and cannot be gamed. Across 54 test cycles, a frontier agent claimed improvement every single time. The oracle showed 56 percent of those cycles had zero or negative real-world impact. Letting the agent's self-verdict gate the loop degenerated into accepting everything, and the agent's best working state eroded by 19 percent over time.
The more interesting finding: making the judge smarter did not fix it. Even a judge given the full change diff, artifact text, and its own verdict history still approved regressions 44 percent of the time and rejected genuine improvements 38 percent of the time. The researchers had specifically hypothesized a stronger judge would close this gap. It did not, and they say that hypothesis is now rejected. What did work was moving the success signal outside the agent's own transcript entirely, into something independently checkable.
This matters because "let the agent decide when it's done" is the default architecture for a lot of long-running coding agents and autonomous workflows shipping right now. The paper's evidence says that architecture is structurally broken for any goal whose success can't be verified from inside the conversation, no matter how capable the underlying model gets. A stripped-down version that returned only a pass or fail verdict, with no explanatory feedback, performed about as well as the full-feedback version. The fix isn't a better rubric or a chattier judge. It's an evaluator that never reads the agent's own account of its work at all.
Call it a check on the current enthusiasm for fully autonomous coding agents: the bottleneck isn't the model's competence, it's who's allowed to grade the homework.