6 个 agent 在赛道上比 Jev 与单一大模型

Asteri @Asteri_eth原推发布于 2026.10.10

在 X 上打开原推

I LAUNCHED 6 AI AGENTS ON A RACE TRACK IN JEV AND JUST LET THEM DRIVE every car is a different agent from my stack > SCOUT is the research agent > ROUTER decides which job goes where > CODER writes the code > CRITIC argues with everything > SIZER watches the budget and the grey one is a single big LLM trying to do all of that alone every corner is one decision. brake late, push, hold the line, defend. the Jev cars get the track state, pick an answer and keep driving. one call per corner, around 20-30ms, zero text the LLM-only car hits turn 1 and starts writing a paragraph about the apex. by turn 2 it's hundreds of tokens deep and the field is gone the others keep lapping it. it keeps thinking my favorite part is CRITIC. it goes for an overtake with 0.71 confidence. anything under 0.85 goes to the pit wall, a human says "hold, not here", and it drops back instead of wrecking five small agents with one job each beat one big model doing everything claude wrote the race plan before the lights went out jev takes every corner code holds the wheel be honest, which car is your agent stack right now?

这条是引用着另一条写的：x.com/asteri_eth/status/2104901781725598103