Claude Code's layered agent tree uses Jev for micro-decisions
Opus writes the plan while Haiku and Sonnet agents execute it, and no-thought forks like the next file or whether to retry are settled by Jev in milliseconds so the large models only see real branches.
THIS IS A STRAIGHT-UP F*CKING CHEAT CODE a new Claude Code mode from Anthropic lets Opus write the plan once, then keeps its context clear of the hundred or so agent transcripts a big run piles up. as tricks go, it sits near the top of the list. Opus 5.5 drafts the whole job as a script before anything runs, and a fleet of cheaper agents carries it out. the script holds the plan, so the only thing handed back to Opus is the final answer. that lets jobs grow while Opus token use falls by a wide margin. Haiku 5.5 agents then fan out and read all the files at once, and every fix is patched by a Sonnet 5.5 agent working on an isolated copy of its own. Opus gets pulled in at just three points. before the run it writes the workflow. when the run ends it reads the verified results, and nothing ships until it has signed off. JEV engineering applies the same idea a level down. decisions that need no thinking (which file comes next, which tool to call, whether to retry or stop) get settled by JEV within milliseconds, so the large models are shown only the forks that branch. launch with: claude --model opus --effort ultracode the whole setup as a tree: Opus 5.5: main session, drafts the workflow ├─ scan: Haiku 5.5 goes through every file ├─ edit: Sonnet 5.5 applies each fix inside an isolated copy ├─ verify: Sonnet 5.5 re-examines every finding ├─ Opus 5.5: sees only the findings that passed verification └─ JEV: resolves micro-forks in milliseconds give Claude Code that tree along with the prompt below. you'll end up running every large job through it. "Reorganize my Claude Code configuration around the tree above. 1. Opus 5.5 is the main session model, ultracode enabled. 2. In each workflow, read-only scan stages use Haiku 5.5, edit stages use Sonnet 5.5, and every edit is made on a separate, isolated copy. 3. CLAUDE.md gets a single new rule: a verifier agent that played no part in a finding must confirm it before it is reported, and a fix loop halts after two rounds without progress. 4. Send quick, deterministic decisions (which tool, which path, whether to retry) through JEV. First show every proposed change as a diff. Leave all files untouched until I give the go-ahead." link in comments