Anthropic has officially added an advisor tool to Claude Code's documentation. Shortly after, mirku posted a tweet sharing a radical layered configuration setup: Opus 5.5 as advisor, Sonnet 5.5 as the main worker, and Haiku 5.5 for routine legwork.

The official documentation is available at code.claude.com/docs/en/advisor. This tool lets Claude ask a more powerful model to review work at critical task checkpoints.

Here's the role division outlined in mirku's tweet:

- Sonnet 5.5: Main session, high-effort mode, writes diffs and runs test suites

- Haiku 5.5: Sub-agent, scans repositories in parallel at 340 tok/s for file discovery and spec document lookup

- Opus 5.5: Background advisor, only intervenes at three specific checkpoints

The three intervention points:

1. Before finalizing a plan: Check if the solution missed authentication invariants or schema contracts

2. After two consecutive test failures: Determine if you're fixing the root cause or falling into a recursive rabbit hole

3. Before marking the task complete: Check if the diff introduces hidden regressions or violates pre-flight rules

Original configuration command shared:

```bash

claude --advisor opus --subagents haiku

```

Commenters pointed out the `--subagents` parameter does not exist. The actual usage per official documentation is:

```bash

claude --advisor opus

```

Or add it to your settings.json:

```json

{

"advisorModel": "opus"

}

```

You can also switch advisors mid-session by entering `/advisor opus` directly.

The video demonstrates mirku's agent tree interface, which clearly shows model role assignments and scheduling status.

mirku also notes that the JEV micro-fork layer uses the same pattern at the next level: over 1,500 mechanical routing branches are resolved within 16ms. Mechanical decisions (which file to open, which tool to call, retry or abort) don't require reasoning, so cutting-edge models only wake up when the execution path actually branches.

Connecting this to mirku's earlier piece on [Harness Engineering](https://x.com/i/article/2098825381931111317), the logic is consistent. Editing prompts only fine-tunes tone, but reworking the harness improves the whole system. As Dario Amodei said when Claude Code launched: you need an interface, a harness, to use models effectively. Models are the CPU, context is the RAM, and the outer layer — file system, permissions, testing, state management — determines whether the model can produce stable output.

The value of the advisor tool is that you don't need to run Opus for the entire session; you only need it to take a look at critical decision points. Sonnet does the work, Haiku handles the legwork, Opus does the final check. Costs go down, while quality stays the same.

The official documentation also mentions several limitations: the advisor requires Anthropic API and does not support third parties like Bedrock; Fable models can also act as advisors but require separate authorization. Additionally, the advisor model cannot be weaker than the main model. For example, if you run Opus 5.5 as your main session, you can configure Opus 5.5 or Fable as advisor, but not Haiku 5.5. See the documentation for the full compatibility table.

When asked if this works in VS Code or the desktop app, the documentation notes that advisor is a feature of Claude Code for terminal, but it is also available in Agent SDK and non-interactive mode, requiring version v2.1.260 or higher.

Layered model setups are not new, but the fact that Anthropic has officially released it as a tool confirms that this approach works. The next question is no longer which model is the strongest, but which work should be assigned to which model.