The strongest model shouldn't write code—it should be the advisor
Anthropic added an advisor tool to Claude Code. Write code with Sonnet, look up documentation with Haiku, and have Opus make decisions only at critical moments. Some people use the same approach for non-code work. The key to AI agents is how you divide responsibilities and where you set human checkpoints.
Anthropic added a new tool to Claude Code called advisor. The poster xsmryze put it bluntly: THIS IS ACTUAL F*CKING GOLD. Indeed, this is almost exactly what every heavy Claude Code user would want.
Core logic is simple: stop making the strongest model do all the work. Day-to-day reading and writing, searching, and coding go to cheaper models. The strongest model only makes key decisions.
Startup:
claude --model sonnet --advisor opus
With this setup:
- Sonnet 5.5 runs the main session and writes code
- Haiku 5.5 explores files and checks API docs in parallel
- Opus 5.5 acts as advisor and steps in only at critical moments
Opus shows up at three moments: before a plan is locked, after the same error appears twice, and before a task is wrapped up. It receives the full session, gives recommendations, and then goes back to the background. You no longer pay Opus prices for every routine operation.
xsmryze also mentions another layer: JEV routing. Which file to choose, which tool to use, whether to retry—these highly deterministic micro-decisions are handled by JEV in milliseconds. The large model only looks at forks that genuinely require thought. This is the same logic decomposed one level further: be fast where you can, think only when you should.
Here's a ready-to-use prompt he provides to have Claude Code rebuild its configuration around this structure:
Rebuild my Claude Code setup around this tree:
1. Run the main session on Sonnet 5.5 with Opus 5.5 as advisor.
2. Create three read-only Haiku 5.5 subagents: search (maps the repo), grep (finds call sites), docs (pulls the API reference).
3. Add one rule to CLAUDE.md: consult the advisor before locking a plan, after the same error twice, and before calling a task done.
4. Route fast deterministic forks (tool choice, path selection, retries) through JEV.
Show me every change as a diff first. No edits until I say go.
Official docs are at code.claude.com/docs/en/advisor. It's currently an experimental feature and requires the Anthropic API or a specific cloud platform.
This approach isn't just for code. N01ennn wrote a long post about how he handed the day shift to an AI employee, Viktor. The day shift means all non-code work: replying to support, triaging bugs, writing changelogs, and producing weekly reports.
He used to be interrupted over and over by the same questions: How do I reset a key? Why is the export empty? Is this a bug, or did I do something wrong? The answers were all in the docs—he was the search bar.
Eventually, he split the work into four loops, each with five parts: trigger, context, draft, gate, and report. Viktor can only draft, not send. Every user-facing action waits for human approval. This gate is the core of the entire system.
Results: 12 replies approved, 3 bug reports turned into issues, the changelog started shipping weekly, one Monday-morning message replaced three dashboards, and 5 hours were saved.
N01ennn put it very accurately: handle support is not a job. It has no start, no finish line, no place to stop. A loop counts. Give AI a loop, and it can do work; give AI a vague task, and it will flail around.
Claude Code's advisor and Viktor's loop share the same core: AI executes, humans decide. The difference is the granularity of the division. The former divides at the model layer—who writes code, who looks up information, who makes judgments. The latter divides at the task layer—who handles support, who organizes bugs, who writes the weekly report.
This offers a lesson: the effectiveness of an AI agent depends more on the lanes you set for it and where you place checkpoints. How strong a single model is matters less.
发布时间: 2026-11-21 11:36