The Advisor Strategy
Article: The advisor strategy: Give agents an intelligence boost by Anthropic
Summary
Anthropic introduces the advisor strategy — a cost-effective pattern for building AI agents that pairs Claude Opus as an advisor with a smaller model like Sonnet or Haiku as the executor. The smaller model runs the task end-to-end and calls on Opus only for the hard decisions, giving near Opus-level intelligence at close to Sonnet-level cost.
The executor calls tools, reads results, and iterates toward a solution on its own. When it hits a decision it can’t reasonably solve, it consults Opus, which reads the shared context and returns a plan, a correction, or a stop signal — but never calls tools or produces user-facing output. This inverts the usual sub-agent pattern, where a large orchestrator decomposes work for smaller workers: here the small model drives and escalates, with no decomposition, worker pool, or orchestration logic.
The results are significant. Sonnet with an Opus advisor scored 2.7 percentage points higher on SWE-bench Multilingual than Sonnet alone while costing 11.9% less per task. Haiku with an Opus advisor more than doubled its BrowseComp score (41.2% versus 19.7% solo) — still below Sonnet alone, but at 85% lower cost per task.
Anthropic has operationalised this as the advisor tool, in beta on the Claude Platform. Developers declare it as a tool in a Messages API request, and the handoff happens inside a single request — no separate round-trips or context management. The executor decides when to consult, and max_uses caps how often. Advisor tokens are billed at the advisor model’s rates, but since the advisor typically returns only a short plan, total cost stays well below running Opus end-to-end.