Layered agent loop with Opus 5.5, Sonnet 5.5 and Jev
Opus acts as the control layer, Sonnet as the builder layer and Jev as the decision layer, using typed answers to choose branches, retries and whether to touch production.
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
This post quotes another post: x.com/loofyb0i/status/2102452178384273431