The advisor
A stronger second opinion at decision points, and its one hard rule.
The advisor
Not a subagent. A server-side tool. Claude calls it when it matters — before committing to an approach, when an error keeps recurring, before declaring a task done. It reads the entire conversation and answers.
lead agent ─── working ──────────────────────────▶
│ │
│ stuck / deciding │ about to call it done
▼ ▼
[ ADVISOR ] [ ADVISOR ]
server-side · sees the whole transcript · you don't invoke it
Turn it on
claude --advisor opus # one session
/advisor opus # mid-session, saved as your default
/advisor off # stop
{ "advisorModel": "opus" }
The hard rule
The advisor must be at least as capable as the main model.
| Main model | Accepted advisors |
|---|---|
| Opus 4.7 or later | Fable, Opus 4.7+ |
| Sonnet 5 | Fable, Opus, Sonnet 5 |
| Haiku 4.5 | Fable, Opus, Sonnet |
On an Opus 5 main, a Sonnet advisor is rejected. Fable or another Opus, nothing else. Haiku can call an advisor but can never be one.
Subagents inherit the advisor and get checked the same way, so the Sonnet Librarian and the Haiku Historian both reach the Opus advisor.
Anthropic API only — not on Bedrock, Vertex, or Foundry. Experimental. Each call costs advisor-model tokens on top of the main model.