Opus 5.5 + Sonnet 5.5 + Jev 的分层 agent 循环
把 Opus 当控制层、Sonnet 当执行层、Jev 当决策层，用类型化答案决定分支、重试与是否碰生产。
holy sh*t, this is the agent setup nobody is talking about yet opus 5.5 + sonnet 5.5 + jev, wired into one loop most multi-model stacks do one of two things: dump everything into one giant context, or hardcode a loop that dies after 5 steps you re-pay for the same history every turn, and your most expensive model ends up babysitting tiny decisions this setup splits the work into layers: > opus 5.5 - control. writes the plan, never the patch > sonnet 5.5 - 3 builders, one branch each, fresh context, sandboxed > jev - decision layer. which branch, which tier, retry or stop, does this touch prod. typed answer in 70–500 ms, ~$0.00003 a call > verifier - isolated. sees the diff and hidden tests, never the builder's reasoning > memory - context resets every lap, history lives in a graph, an offline pass turns wins into rules opus sets the direction. sonnet tests it in code. jev makes every small call in between and never writes a token the shift: your frontier model stops being your router
这条是引用着另一条写的：x.com/loofyb0i/status/2102452178384273431