Lastly, you can now give a Claude Managed Agent an advisor: a stronger model it

An advisor lets one Claude model keep working on a Claude Managed Agents task while consulting a more capable model at chosen moments. Anthropic's August 7 announcement describes it as a stronger model the working agent can call mid-session for a second opinion. Its example gives an API-design agent running Claude Sonnet 5 a Claude Opus 5 advisor, with a description saying the agent escalates irreversible decisions to it.

Anthropic's multiagent documentation describes how a consultation works. Only the session's primary thread can consult the advisor; subagents cannot. Each consultation sends the advisor the primary thread's conversation so far, and the advice comes back as a message while the working agent keeps doing the work itself. The advisor model must be at least as capable as the agent's own model, and a failed consultation does not fail the agent's turn.

The cost grows with the session. Consultations are billed at the advisor model's rates and count against any session budget, and because each one carries the whole conversation so far, late consultations in a long session use more input tokens. The announcement reports no measured effect on task results. The setup has also moved since August: the announcement added the advisor as an entry in a `coordinator` roster, while the current documentation's examples use a separate `advisor` setting in the agent's `multiagent` block and say the API still accepts the older type.