Claude Code advisor mode with Jev taking the small decisions first

rvaniaaa @rvaniaaaaOriginally posted 2026.10.28

Open original on X

I still don't understand why you're overpaying for Opus when Anthropic's own docs already show the fix. Three models share one task, and each one only does the part it is cheapest at. Start Claude Code like this: claude -- model sonnet -- advisor opus Now you stop paying Opus prices for work that never needed Opus: → Sonnet runs the session and edits the code → Haiku 5.5 searches the repo and docs for up to 40x less per token than Opus 5.5 → Opus sits in the background as the advisor and reads the whole session, tool calls included Claude typically calls Opus at three moments: → before it commits to a plan → when the same error comes back → right before it says the task is done Sonnet drives, Haiku scouts and Opus signs off. One level lower, most of an agent's day is tiny choices, like which file to open, which tool to run, or whether to retry. A decision model like Jev takes those first, so Sonnet and Opus only see the moments that are really open. Here is the full setup: Sonnet runs the main session Explore on Haiku 5.5 does the searching and doc lookups, read-only Opus stays on call as the advisor Give Claude Code this prompt and it sets everything up: "Set up my Claude Code this way: Main model Sonnet, advisor Opus, saved as advisorModel in my settings. One read-only subagent named exactly Explore on claude-haiku-5-5, so it replaces the built-in one. One line in CLAUDE.md: consult the advisor before locking a plan, after the same error twice, and before calling a task done, and no more often than that. Check what could switch the advisor off, such as CLAUDE_CODE_DISABLE_ADVISOR_TOOL and DISABLE_TELEMETRY, and only report it. Show every change as a diff first, and edit nothing until I say go." Docs: https://code.claude.com/docs/en/advisor Four things to know before you turn it on: → every advisor call re-reads the whole transcript with no cache, so on a 200k session that is about $0.80 of input against $0.04 for a normal request, which means a few calls are enough → the built-in Explore agent runs on your main model, and only a subagent named exactly Explore replaces it, because any other name does nothing → turning the advisor on or off keeps your prompt cache, while switching models rebuilds it → it needs the Anthropic API, and it stays off if telemetry is disabled

This post quotes another post: x.com/rvaniaaaa/status/2109385796751114346