推文

@harden_369741 · 2026-10-12 01:16

Your 10-agent swarm is just a very expensive way to get one agent's output. Vals AI ran the numbers on Vibe Code Bench: multi-agent teams burn 1.8x to 5.1x the tokens of a single agent. Out of four matchups, exactly one showed a statistically significant quality gain. One. I learned this the dumb way. I once pointed three agents at the same repo and they spent the whole afternoon overwriting each other's fixes while I sat there playing referee. More agents, more meetings. Brooks's law didn't die, it just got a GPU upgrade: adding manpower to a late project makes it later. Turns out the same applies when the manpower is artificial. The bottleneck was never model intelligence. It's coordination. It always was. #AI #agents #CodingAgents

曝光 5 · 评论 0 · 点赞 0 · 书签 0 · 曝光/时 3.2787343044404564