Claude Code 分层：Jev 管小决策，Opus 只做规划
作者把 Opus 5.5 从日常执行里撤出，让 Jev 负责选工具、查文件等小决策，Sonnet 干活、Fable 当顾问，一次跑测出 2 个被跳过的测试。
I DELIBERATELY STOPPED OPUS 5.5 FROM DOING MOST OF THE WORK. AND THE STACK GOT BETTER. I wanted to see what would happen if I stopped using the strongest model as a universal employee inside Claude Code. I left Opus 5.5 with the main session, planning, hard decisions, and final review. Sonnet 5.5 does the regular work: writes code, runs tests, reads the codebase, and looks for docs. Fable 5.1 sits next to it as an advisor. it sees the whole session, but stays quiet until it actually notices a problem. and Jev sits between them. it handles small decisions like choosing a tool, checking a file, or deciding whether to keep going, so Opus doesn't have to wake up for every single step. one run showed exactly why I wanted this setup. the worker finished the task. the tests were green. Opus was already about to close the work. Fable noticed 2 skipped tests. the task went back, Sonnet finished the missing work, it ended with 50 passed, and only after that did Opus send the final result. > Claude Code - the whole workflow > Opus 5.5 - planning + final review > Sonnet 5.5 - coding + tests + research > Fable 5.1 - advisor > Jev - routing + small decisions and this is more interesting to me right now than another benchmark. the question is no longer just which model is smarter. the question is why the most expensive model should see every decision inside your agent stack at all. the video shows the whole run. in the article, I went deeper into what happens when you start removing big models from the places where they were never needed in the first place.
这条是引用着另一条写的：x.com/model_culture/status/2105354309680435332