Dots 用多智能体锦标赛替代单智能体
作者实测 Dots 让 5 个廉价智能体并行做同一任务、由独立评审挑一个晋级，90 秒完成单智能体 11 分钟仍失败的任务。
Dots just turned a swarm of agents into something that actually finishes work, and the trick is that most of them are allowed to fail Most multi-agent setups fall apart because every agent is trusted equally, so one bad output quietly poisons the whole run. Dots runs them as a tournament instead, the same task goes to several agents and only the result that survives review moves on: → Fan out: one task goes to 5 agents at once, each a different approach → Review: a separate agent scores all 5 against the spec, not the vibe → Keep one: the winner advances, the other 4 are thrown away on purpose → Repeat: the next step fans out again from the survivor • Why throwing away work wins > A wrong answer never advances, so errors can’t compound downstream > Five cheap attempts beat one expensive agent trying to be perfect > The review agent never built anything, so it has no result to defend • What the swarm actually costs > 5 parallel drafts run on the cheap tier, not the frontier one > Only the review step touches the expensive model > A full fan-out-and-judge round lands near $0.08 And the part that broke my brain: One careful agent took 11 minutes and still missed the spec. Five throwaway agents plus a judge nailed it in 90 seconds. Same task. Wasting four results was faster than trusting one. The metric for a swarm isn’t how many agents you run. It’s how many you’re willing to discard ↓
这条是引用着另一条写的：x.com/fleyta88/status/2105656875387625507