WHAT NEW IN X10.30 FRI

AI collections from X · 22 collections · 61,986 posts

jevVIDEO

Claude Code swarm: one Haiku agent per file, Jev routes model choice

Sonnet 5.5 leads 8 parallel slots while Haiku handles one file per agent, and Jev resolves 4,868 forks in code over model, slot and retry, escalating only genuinely split ones to Sonnet.

VETRA
@0X_VetraArchived 12 days agoOriginally posted 2026.10.17
VETRA @0X_VetraOriginally posted 2026.10.17
Open original on X

THIS IS ACTUAL F*CKING GOLD Anthropic just shared one of the best tricks for Claude Code. The idea: one prompt, one lead, and a swarm of Haiku agents doing one file each. 24 files converted to TypeScript for $0.71 in 3 minutes. Opus alone was 22% done at $1.05 and still running. You just launch Claude Code like this: claude --swarm haiku --lead sonnet here's the split: → you type one prompt and never touch it again → Sonnet 5.5 splits the job and opens 8 parallel slots → 21 files go to Haiku 5.5, one file per agent, ~2k tokens each → the 3 hardest files route to Sonnet 5.5 instead that is the whole trick: one file per agent means tiny context, and tiny context means the cheap model is enough. Sonnet only comes back at the end to merge 24 diffs into one and run the suite. 148/148 green. $0.029 per file. zero follow-up prompts. Sonnet leads. Haiku does the volume. You just review the diff. JEV engineering is the same move one layer down: 4,868 forks this run - which model, which slot, retry or stop - resolved in code, and only the genuinely split ones go up to Sonnet the full tree > Sonnet 5.5 leads the session on medium effort and splits the work > swarm runs 24 agents across 8 parallel slots > each agent gets exactly one file and its own context > JEV sends sharp forks to Haiku 5.5 and split ones to Sonnet 5.5 > Sonnet 5.5 merges 24 diffs into one and runs the test suite > you review a single diff paste the tree and this prompt into Claude Code ↓ "Rebuild my Claude Code setup around this tree: 1. Run the session with Sonnet 5.5 as lead and a Haiku 5.5 swarm, 8 parallel slots. 2. Split file-level work so each agent gets exactly one file and its own context. 3. Route per-file model choice through JEV: sharp forks go to Haiku, split ones escalate to Sonnet. 4. Add one rule to CLAUDE.md: the lead merges all diffs into one and runs the full suite before anything is reported done. 5. No follow-up prompts - the run finishes or stops, it never asks me mid-way. Show me every change as a diff first. No edits until I say go." ↳ http://code.claude.com/docs/en/model-config save it, you'll use this every single day.

This post quotes another post: x.com/cyrilxbt/status/2110112722423816211

x.com/0x_vetra/status/2111385535511949502Archived 2026.10.18 10:03