Critics loops vs 0 shot
热度趋势
趋势数据积累中
百分比基于当前可用热度信号，而非评论数或独立用户人数。
一个Three.JS项目是使用Claude 5.0 Opus完成的，动用了18个代理，每个代理负责非常具体的任务。开发者表示，这是一个“0 shot”作品，仅需一次小的修正提示来解决建筑物之间的间隙和一条被阻挡的路径。该项目没有使用任何批评循环，开发者推测，如果只用3-4个代理来完成，结果会差很多，即使有QA代理的建议，也可能不如当前的效果。
Yes this is all pure Three.JS, 0 external assets. it was mostly a 0 shot work, but i did do 1 small corrections prompt (it left some big gaps between buildings and one path was blocked).
I am wondering this: Is it possible people are wasting time and money in "critics loops" when really, agents is all you need.
This was done with 18 agents all working on very specific tasks. 0 critics loops. If Claude tried to do this with just 3-4 agents, i would guess it produces a much worst results, and then the QA agent would give some imperfect recommandations and it would probably not be as good as what i've gotten here.
But most importantly, the QA agent will never beat an actual human tester. So it feels much more optimal to me to get the human to do the critics loop.
The other issue is, 18 agents for 24 hours would burn anybody's token budget lol