AI Agents Spend $3,000 in 6 Days to Complete Two Papers, Both Rejected, Exposing Judgment Shortcomings

TechFlow News, August 1, AI agents completed two studies and produced papers within 6 days and a $3,000 budget, but both were rejected by reviewers. The main experiment used Claude Opus 4.8 paired with the OpenClaw framework; the agent ran hundreds of experiments, debugged GPUs, and completed LaTeX typesetting, but when facing negative reviews, it only narrowed conclusions and added limitations, rather than redesigning the experiments. At the end of both rounds of experiments, more than half of the budget remained, exposing a problem of lack of creativity rather than insufficient funds. This case shows that the bottleneck of current AI agents lies in judgment and creativity, rather than execution capability.