Claude Advisor tool - Pair Opus as advisor with Sonnet or Haiku as executor

Claude Advisor Tool lets your Sonnet or Haiku agent consult Opus mid-run when it hits a decision it can't resolve on its own.

It solves the core tension in production agent work: running Opus on every step is expensive, but running Sonnet or Haiku alone means weaker outcomes at hard decision points.

The advisor strategy inverts the typical multi-agent pattern. Instead of a large orchestrator delegating down to smaller workers, a smaller model drives the full task and escalates up to Opus only when it needs guidance. Opus reads the shared context, returns a plan or correction, and the executor resumes. The advisor never calls tools or produces output directly.

Add it with one tool declaration in your existing Messages API call. The full handoff happens inside a single /v1/messages request, no extra round-trips or orchestration logic.

Key results from Anthropic's evals:

• Sonnet + Opus advisor: +2.7pts on SWE-bench Multilingual, 11.9% lower cost per task vs Sonnet alone

• Haiku + Opus advisor: 41.2% on BrowseComp vs 19.7% solo, at 85% lower cost than Sonnet solo

Key features:

• One-line change to your existing Messages API call

• Full advisor loop inside a single request

• max_uses cap for cost control per request

• Advisor tokens billed separately at Opus rates

Available now in beta on the Claude Platform.

Perfect for API developers and engineering teams building coding agents, browser automation, and long-horizon tool-use workflows where both task quality and token cost are tracked metrics.

At what point in your agent runs does the executor tend to break down reasoning failures, ambiguous tool results, or something else?

P.S. I hunt the latest and greatest launches in tech, SaaS and AI, follow to be notified → @rohanrecommends

This is something most of the products have built the systems of creating kind of Manager agents using higher FLMs and run the executions in the lower parameters LLMs.

Just curious to know what innovation is here. For me it looks more of a workflow orchestration of agents.

Interesting setup — separating advisor and executor roles makes sense. Did this structure come from real workflow issues or was it designed from the start?

This was the day of the year that PH got put to use; we already implementedthis into a social engagement pocess; it can be dumb most of the time, intelligent only when needed. (Like my wife.)

this feels like giving your agent a “call senior engineer when stuck” button 😅

except now the senior only shows up when things are already on fire

actually this is super cool, but feels like all the “real” agent features are slowly moving to API-only 👀

any plans to bring something like this to the subscription side, or is that intentionally kept separate?

From my experience running multi-agent workflows, executors mostly break down on ambiguous tool results. The model picks wrong data from a dump or misinterprets a failure mode. Routing those moments to Opus for a second opinion sounds like a clean fix. I call this supervisor mode in my agents.

This is a smart pattern for production agent systems. I've been running Sonnet as the main loop for tool-heavy workflows and the cost adds up fast when you need Opus-level reasoning on edge cases. Having the advisor as a on-demand escalation path instead of running Opus on every turn is exactly the right trade-off. Curious how the handoff latency feels in practice — does the shared context approach keep it under a second?

I have a question: Are these Claude tools helpful for a non software engineer?

Nice! Could have been helpful when I was burning through Opus like it was going out of style trying to build my mobile apps.