Rohan Paul (@rohanpaul_ai)T2·Agents
71 pts
A joint UCLA and AWS paper had Claude Opus 5 agents solve polynomial systems with hidden shortcuts, finding the agent solved 103 of 108 systems within 600 seconds versus 60 for plain msolve, yet was faster on only 12 of those 60. The study concludes that the thinking time agents spend spotting shortcuts often cancels the time saved, so agents should target problems standard solvers cannot finish, and reusable solvers only pay off when enough similar problems justify the 11.4-hour setup.