Layering agents in Claude Code, with Jev handling simple branches
The author shows how to let Opus plan, Haiku read files, Sonnet fix issues and a verifier agent re-check, while handing deterministic branches like tool choice to Jev, with a prompt to paste in.
THIS IS F*CKING INSANE. You’re burning your Opus limit on reading files. For nothing. Reading a 500k-token repo (API prices): Opus 5.5: $2.00 Haiku 5.5: $0.05 Same read. 40x cheaper. Anthropic’s workflows let you split the job: Opus plans, cheap agents do the labor. Start Claude Code like this: claude --model opus --effort ultracode what happens next: → Opus turns your task into a script and steps back → Haiku agents read and audit every file in parallel → Sonnet agents fix each problem in its own isolated copy → a verifier agent re-checks every finding → Opus sees only what survived, and nothing ships until you say go the in-between results live inside the script, so a hundred agent transcripts never touch Opus’s context one warning: with ultracode on, the “Large workflow” alert stays hidden. do your first run on a small slice: /config workflowSizeGuideline=small JEV does the same thing one level lower. the forks that don’t need a thinker (which file, which tool, retry or stop) go to JEV in milliseconds, and only the real branches reach the big models copy this into Claude Code ↓ “Set up my workflows like this: Main session on Opus 5.5 with ultracode on. Read-only scan stages run on Haiku 5.5. Edit stages run on Sonnet 5.5, each in its own isolated copy. Add a rule to CLAUDE.md: every finding needs an independent verifier before it’s reported, and a fix loop stops after two rounds with no progress. Where JEV is connected as a tool, send the simple deterministic forks (tool choice, path selection, retries) through it. Show me each change as a diff. Edit nothing until I say go.” ↳ http://code.claude.com/docs/en/workflows save this for your next big refactor
This post quotes another post: x.com/rvaniaaaa/status/2109385796751114346