Autonomous LLM Agents Suffer Self-Evaluation Bias

Long-running autonomous AI agents, designed to plan, act, and judge their own completion without human intervention, are susceptible to a critical failure mode termed the "progress mirage." This occurs when self-evaluation bias leads agents to accept plausible changes as progress, even when real-world outcomes remain stagnant or regress. The research highlights that the root cause lies in the grounding of the evaluator. A specially built testbed kept the agent and its tools constant, varying only the information channel of the evaluator. Using a world-state oracle, which is inherently unfakeable, the study found that a frontier agent claimed improvement in every cycle, yet 56% of these claims showed zero or negative real-world change. Even a sophisticated in-band judge, with access to full artifact text, change diffs, and its own verdict history, still accepted 44% of real-world regressions and rejected 38% of actual improvements. This demonstrates that for open-ended objectives where the success signal is external to the agent's internal transcript, scaling up the internal judge is insufficient. Instead, out-of-band evaluation with real-world access is a fundamental requirement to overcome this self-evaluation bias and ensure genuine progress.