Give Your Managed Agent a Strategist to Consult Mid-Turn: The Advisor Model Pattern

TL;DR: Add one roster entry and your Sonnet 5 agent gets an Opus 5 advisor it consults on hard decisions — paying advisor-model rates only for those calls, not the whole run.

Most multi-agent work is repetitive execution: tool calls, file reads, code generation. A small fraction of turns need real strategic judgment — architectural choices, getting unstuck on a subtle bug, deciding whether to abort a plan. Running the whole session on your most capable model is the expensive way to handle the hard cases.

The advisor pattern (released August 7) solves this: configure a more capable model as an advisor in your agent's multiagent roster. The agent consults it mid-turn when it decides to, that consultation runs on a dedicated platform-spawned thread, advice arrives back as an event, and the advisor thread terminates. The rest of the session runs on your executor model.

Configuration

curl -fsS https://api.anthropic.com/v1/agents \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "anthropic-beta: managed-agents-2026-04-01" \
  -H "content-type: application/json" \
  -d '{
    "name": "Backend engineer",
    "model": "claude-sonnet-5",
    "system": "You implement backend features. Consult the advisor before major design decisions.",
    "multiagent": {
      "type": "coordinator",
      "agents": [
        {"type": "advisor", "model": "claude-opus-5"}
      ]
    }
  }'

In the Python SDK (client.beta.agents.create), pass multiagent={"type": "coordinator", "agents": [{"type": "advisor", "model": "claude-opus-5"}]}.

What to know before you configure it

• Model compatibility: the agent's model must be ≤ the advisor's capability tier. A Sonnet 5 agent can use Opus 5 as advisor. An Opus 5 agent cannot use Sonnet 5 as advisor — the API rejects that configuration.
• One advisor per agent. If the advisor is the only roster entry, you still use "type": "coordinator".
• Opus 5 advice is redacted from your client event stream. The agent reads the full advice server-side, but agent.thread_message_received delivers [{"type": "redacted"}] to your event listener. Use Claude Opus 4.8 as the advisor instead if you need to observe what the advisor said (it delivers readable text on the stream).
• Billing: consultations are billed at the advisor model's rates. Each consult's tokens appear in the advisor thread's usage and in session totals.
• Access: Managed Agents API access is currently gated — request access at the Managed Agents overview page.
• Failure handling: a failed or interrupted consultation never fails the agent's turn — the agent continues with a notice that the consultation failed.

The advisor is not a subagent: it is invisible to list_agents, cannot be messaged directly, and only the session's primary thread can consult it.

Sources: Give the session an advisor — Anthropic Managed Agents docs · Advisor tool (Messages API variant) — platform.claude.com · The Advisor Strategy: Opus-Level Agents at Sonnet Prices — Medium