The advisor strategy introduces a new way to enhance the intelligence of AI agents by pairing Opus as an advisor with Sonnet or Haiku as executors. This approach allows developers to achieve near Opus-level intelligence while maintaining lower operational costs.
With the newly launched advisor tool on the Claude Platform, making this integration is as simple as a single line change in the API call. The executor, whether Sonnet or Haiku, manages tasks from start to finish, consulting Opus only when it encounters complex decisions that require additional guidance.
How It Works
In this setup, the executor performs the task and, when necessary, invokes Opus for advice. Opus analyzes the shared context and provides a plan or corrective action, allowing the executor to continue without generating user-facing outputs itself.
This method contrasts with traditional models where larger systems decompose tasks and delegate them to smaller units. Instead, the advisor strategy allows a more cost-effective model to drive the process, applying advanced reasoning only when essential.
Performance Benefits
Evaluations indicate that using Sonnet with Opus as an advisor results in a 2.7 percentage point improvement on the SWE-bench Multilingual1 benchmark compared to Sonnet alone, while also reducing costs per task by 11.9%.
When Haiku operates with Opus as an advisor, it achieved a score of 41.2% on the BrowseComp benchmark, significantly higher than its solo score of 19.7%. Although Haiku's scores trail Sonnet's, it remains a cost-effective alternative, being 85% less expensive per task.
Implementation and Cost Structure
To implement the advisor strategy, developers need to declare advisor_20260301 in their Messages API request. This allows for seamless integration without additional context management. The advisor generates short plans, typically between 400-700 tokens, while the executor handles the complete output at a lower rate, ensuring overall costs remain low.
Next Steps
Developers are encouraged to test the advisor strategy by running evaluations with Sonnet solo, Sonnet with Opus, and Opus alone. This will help in understanding the performance improvements and cost efficiencies gained from this new approach.
Conclusion
The advisor tool is currently available in beta on the Claude Platform, offering a promising solution for developers looking to enhance the intelligence of their AI agents without incurring high costs.
