Claude Code subagent split with JEV handling micro-decisions
Opus plans, Haiku explores, Sonnet builds, Fable audits, and the author says JEV absorbs the no-thinker micro-decisions to spare the big models.
THIS IS ACTUAL F*CKING GOLD Anthropic just documented the single best subagent setup for Claude Code The setup: let Opus 5.5 run the main session, delegate sweeps to Haiku 5.5 at 340 tok/s, run Sonnet 5.5 builder subagents, and keep Fable 5.1 on call as the dedicated adversary: Side tasks in isolated context windows with zero Opus bloat: each subagent runs independently and Opus only receives the verified summary diff You just launch Claude Code like this: claude --model opus here's how it splits: → Opus 5.5 runs the main session, plans architecture, and delegates tasks → Haiku 5.5 powers read-only explorer subagents at 340 tok/s for fast search → Sonnet 5.5 runs builder subagents in isolated contexts for edits and tests → Fable 5.1 acts as adversary on call to audit subagent plans and findings → you stop flooding your main conversation with thousands of raw tool logs Fable 5.1 only steps in at three moments: → before a plan locks: audits subagent task decomposition and tool scopes → when a subagent hits the same error twice: stops symptom patching and redirects → before calling it done: cold adversarial review of diffs before Opus sign-off Opus orchestrates · Haiku explores · Sonnet builds · Fable audits 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 > Opus 5.5 runs the main session and plans architecture > explorer subagent maps files and call sites on Haiku 5.5 at 340 tok/s > builder subagent patches code and runs tests on Sonnet 5.5 > Fable 5.1 stays on call as adversary to audit plans, errors, and PR diffs > subagent contexts stay isolated so main session context stays lean > 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 main session on Opus 5.5 with full reasoning depth 2. Define custom subagents in .claude/agents/: explorer on Haiku 5.5 (read-only search at 340 tok/s) and builder on Sonnet 5.5 (isolated edits and tests) 3. Keep Fable 5.1 on call as dedicated adversary to audit subagent plans, double-failure loops, and PR diffs 4. Add one rule to CLAUDE.md: consult Fable before locking architectural plans, after two subagent errors, and before final PR sign-off 5. Route fast deterministic forks through JEV, show all changes as diffs first. No edits until I say go" ↳ http://code.claude.com/docs/en/sub-agents save it, you'll use this every single day
This post quotes another post: x.com/mirku21/status/2098836008468992471