Jev + Opus 5.5 二十秒分流一万条任务

Dumble @Y0urDumble原推发布于 2026.10.11

在 X 上打开原推

this is f*cking criminal I made Jev + Opus 5.5 untangle 10,000 tasks in 20 seconds, and it completely breaks the "one model does everything" setup most developers throw their best model at the whole backlog. every typo, every rename, every one-line config fix goes through Opus. you burn frontier tokens on work that needed a yes/no, and when you add more agents to go faster, two of them open the same file and you're stuck merging two confident versions by hand this setup kills the single-model bottleneck and splits the stack: > triage: Jev reads all 10,000 tasks once and returns one typed decision per task (trivial / routine / hard) in ~90ms > the comb: Jev closes everything it can on its own. 9,214 tasks never touch Opus > 8 isolated lanes: refactor, tests, bugfix, docs, migrate, perf, security, ui. one helper owns one lane > zero shared files: no lane ever touches another lane's work, so nothing gets rewritten twice > Opus 5.5 only gets the hard 786: it works its lane and ships 41,096 lines the numbers embarrass the usual agent setup: > 12.7x fewer Opus calls than sending everything to Opus > $0.38 total Jev spend for the entire triage > 0 collisions across 8 parallel lanes > 10,000 tasks, 1 knot, 20 seconds Jev decides fast. Opus does the work you stop throwing your best model at everything. you let each model own the one job it's best at

这条是引用着另一条写的：x.com/0xcodila/status/2107119130239185232