Advisor: let a working agent consult a stronger model mid-turn

Why it matters

Running every turn on the strongest tier is safe and expensive. An advisor keeps a cheaper model in charge and buys stronger judgment only for the decisions that need it, at a cost you can measure.

What to practise

In the Claude Managed Agents beta, add an {"type": "advisor", "model": ...} entry to an API-designer agent’s multiagent roster. The advisor tool takes no input, so the system prompt sets the consultation policy: consult before decisions that are expensive to reverse. Give it an API whose money-moving endpoint needs an idempotency scheme. Watch for a thread with agent_name set to anthropic.advisor on the primary stream, and read that thread’s usage.list_cost. Note that a redacted consultation still shows its thread, timing and usage.

How it fits

This is the advisor form of escalation in 2.1.2: the working model keeps the loop and takes guidance; it doesn’t hand over a replacement answer. Use the measured consultation cost against 2.1.1’s single-tier baseline.