Agent stacks have been getting more modular lately, and Anthropic’s latest update leans into that trend with a practical twist. In a new post, “The advisor strategy: Give agents an intelligence boost”, the company introduces an advisor tool for the Claude Platform—designed to make it easier to combine Opus as an advisor with Sonnet or Haiku as an executor inside a single agent run.
The idea is straightforward: keep day-to-day execution on a cheaper model, and only “page in” a more capable model when the agent hits a real decision point.
The advisor strategy, in Anthropic’s framing
Under this strategy, Sonnet or Haiku runs the task end-to-end: calling tools, reading results, and iterating toward an answer. When the executor reaches something it “can’t reasonably solve,” it consults Opus for a plan, correction, or a stop signal—then continues.
A key constraint is architectural: the advisor never calls tools and never produces user-facing output. It only provides guidance based on shared context. That keeps the expensive model’s role narrow and (in theory) predictable.
Anthropic positions this as an inversion of a common sub-agent structure where a large orchestrator decomposes work and delegates to smaller workers. Here, the smaller model stays in the driver’s seat, escalating only when needed—without introducing separate orchestration logic or extra state management.
Benchmarks and cost notes
Anthropic reports that, in its evaluations, Sonnet with Opus as an advisor delivered a 2.7 percentage point increase on SWE-bench Multilingual compared to Sonnet alone, while reducing cost per agentic task by 11.9%.
The post also claims improvements for Sonnet + Opus advisor on BrowseComp and Terminal-Bench 2.0, while costing less per task than Sonnet alone in those evaluations.
For Haiku, the reported effect is more dramatic on BrowseComp: Haiku with an Opus advisor scored 41.2% vs. 19.7% solo. Anthropic adds that Haiku + advisor still trails Sonnet solo by 29% in score, but costs 85% less per task than Sonnet solo.
(As with any benchmark write-up, the footnotes matter: the post specifies various settings such as thinking being turned off, tool availability, effort levels, and which runs used Anthropic’s suggested system prompt.)
The advisor tool: a server-side handoff in one request
The new advisor tool packages this pattern into a one-line change in the Messages API: developers declare advisor_20260301 as a tool, and the executor model decides when to invoke it.
Anthropic says the handoff happens inside a single /v1/messages request—with no extra round-trips or manual context management. When invoked, Anthropic “routes the curated context” to the advisor model, gets back a plan, and continues execution.
A few implementation details called out explicitly:
- Pricing: advisor tokens are billed at the advisor model’s rates; executor tokens at the executor’s rates. Anthropic says the advisor typically generates ~400–700 text tokens of planning output.
- Cost controls:
max_usescan cap advisor calls per request, and advisor tokens are reported separately in the usage block. - Tool compatibility: the advisor tool can sit alongside existing tools like web search and code execution in the same loop.
Availability and how to enable it
Anthropic says the advisor tool is available now in beta on the Claude Platform. Getting started requires:
- Adding the beta feature header:
anthropic-beta: advisor-tool-2026-03-01 - Adding
advisor_20260301to the Messages API request - Modifying the system prompt based on the use case (Anthropic links suggested prompts in the docs)
The company also recommends running an internal eval suite across Sonnet solo, Sonnet executor + Opus advisor, and Opus solo to compare results.
Original source: https://claude.com/blog/the-advisor-strategy
