Did you know even frontier AI agents still struggle to admit when a research project is going nowhere, so they keep wasting time and tokens? Here’s why. Researchers gave AI agents six days, $3,000 in API credits, GPU compute, web access, subagents, and two research questions taken from unpublished NeurIPS submissions. Both agents ran the experiments and wrote complete papers without a human writing or editing the code. The researchers who had spent months studying the same questions scored the papers 2 out of 6 and 1 out of 6: clear rejections. The agents were capable of doing the work. They managed compute, debugged code, analyzed results, and responded to reviewer feedback. What they could not do was recognize when the underlying approach had failed. When early experiments weakened their original ideas, they kept making small adjustments. They narrowed claims, added caveats, and polished results their own reviewers repeatedly said were not strong enough. Their self-reviews identified many of the same flaws the human experts later raised. The agents could see the problems. They could not judge when those problems meant the project needed to be redesigned or abandoned. That is a different failure from hallucination or poor coding. Today’s agents are generally optimized to complete a requested workflow. They are much less reliable at deciding whether continuing that workflow is still the best use of time, compute, and money. Fixing this will require more than better models. Long-running agents need explicit stop conditions, budget and time checkpoints, confidence tracking, and the ability to escalate, replan, or abandon an approach when the evidence turns against it. Operators also need to see how those decisions were made. A polished final paper only shows that the agent finished. The execution history will show whether it recognized a dead end, responded intelligently to criticism, and used its resources well.