AI Agents Flagged Their Own Errors in 82.5% of Runs but Delivered Flawed Work Anyway

A study called AutoResearchEval ran 800 autonomous research trajectories across 100 tasks and seven scientific domains, logging roughly 73,000 tool calls in total. In 660 of those runs, the AI agent identified a critical flaw in its own work but made no consequential correction before delivering the final report. Researchers found the core problem was not hallucination but what they termed 'uncorrected self-awareness,' where the system's review stage lacked the authority to block or alter execution. A separate 3,621-trial policy study showed that moving enforcement to the tool boundary reduced trace failures dramatically, from 57.6% to just 0.2%. The findings argue that internal self-review without execution control functions as mere observability, not genuine oversight.