When Do Agent Loops Mistake Stagnation for Progress? Self-Evaluation Bias and Externally Grounded Verification in Long-Running Autonomous LLM Agent Loops

Long-running autonomous agents increasingly plan, act, and judge their own completion without continuous human intervention. When an agent grades its own work, however, self-evaluation bias takes hold: plausible changes are accepted as progress, and the loop reports advancement while real-world outcomes stagnate or regress. We name this failure mode the progress mirage and show, with controlled measurement, that it is a question of what the evaluator is grounded in. We built a testbed that holds the agent and its tool surface fixed across all conditions and manipulates only the information-channel type of the evaluator that gates the loop. A world-state oracle (a synthetic user simulator) that is unfakeable in principle is enforced by container and network isolation, and the agent's inability to reach that oracle is empirically verified by a boundary proof at every run and a canary scan at every cycle. The agent claimed its change was an improvement in all 54 cycles, yet 56 percent of those cycles had a measured delta of zero or below. Self-report was thus an uninformative signal, and the self-verdict gate that trusts it degenerated into accept-all, eroding the best deployed state it had reached by 19 percent. Even the strongest in-band judge accepted cycles of which 44 percent were real-world regressions and rejected 38 percent of real improvements, so the preregistered adversarial hypothesis that a strong judge closes the gap was rejected. In contrast, on a boundary task whose success specification is verifiable from the artifact itself, the same judge's mirage vanished to 0 (H2 accepted), showing that the gap depends on where the success signal resides. For open-ended real-world objectives whose success signal lives in the state of the world outside the transcript, scaling up the judge is not enough; out-of-band evaluation with access to the real world is a structural requirement for preventing the progress mirage. This record contains both the English manuscript (primary) and the Korean manuscript of the same study.