Claude Code 多模型分工技巧，并把无需思考的分支交给 Jev

mirku @mirku21原推发布于 2026.10.28

在 X 上打开原推

Boris Cherny, Head of Claude Code at Anthropic, just shared an official tip before Fable 5.5 ships The setup: fold routine subagents into Sonnet 5.5, keep Opus 5.5 on call, and only spawn Haiku 5.5 subagents when tasks hit 340 tok/s parallel sweeps Multi-agent sprawl kills execution: passing state through custom routers creates telephone-game context loss, bloated frankenprompts, and 2x latency multipliers You just launch Claude Code like this: claude --model sonnet --advisor opus here's the split: → Sonnet 5.5 runs the folded main loop: writes code, runs tests, and loads skills dynamically without telephone-game context loss → Haiku 5.5 swarms disposable subagents: maps repo trees, greps call sites, and parses API specs at 340 tok/s with zero history bloat → Opus 5.5 stays on call as advisor: ingests full multi-turn session transcripts on high-leverage forks without burning tokens on routine edits → you stop losing intent across subagent boundaries and cut execution latency in half while keeping your primary context lean Opus only steps in at three moments: → right before a plan locks: evaluates architecture and contract invariants → the second an identical test or build error hits twice: breaks debugging loops before tokens burn → right before declaring a complex task complete: audits diff contracts and verify passes before merging Sonnet builds · Haiku swarms · Opus reviews JEV engineering is the same move one layer down: the forks that need no thinker (which file, which tool, retry or stop) go to JEV in milliseconds, and the big models only see the ones that split the full tree > Sonnet 5.5 runs the folded main loop, applies skills on demand, and executes code edits > Haiku 5.5 swarms parallel read-only scouts (search, grep, docs) at 340 tok/s > Opus 5.5 stands on call as advisor, evaluating plans and reviewing diffs > Routine subagents fold into the main loop to kill telephone-game context loss > Subagent payloads pass as plain text with zero conversation history dump > JEV routes micro-forks in milliseconds paste the tree and this prompt into Claude Code ↓ "Rebuild my Claude Code setup around this tree: 1. Run the primary session on Sonnet 5.5 with Opus 5.5 as advisor via --advisor opus 2. Fold single-tool subagents into the main loop using progressive disclosure and dynamic skills 3. Spawn disposable Haiku 5.5 subagents only for parallel file discovery and repo sweeps at 340 tok/s 4. Add one rule to CLAUDE.md: consult the advisor before locking a plan, after repeating the same error twice, and before marking a task done 5. Route fast deterministic forks through JEV, show all changes as diffs first. No edits until I say go" ↳ http://claude.com/resources/articles/building-agents-when-to-fold-subagents-into-your-main-agent-loop bookmark this, stop building multi-agent sprawl when a folded loop runs 2x faster

这条是引用着另一条写的：x.com/mirku21/status/2098836008468992471