Jonathon's AI Wiki

The Advisor Strategy (advisor_20260301)

5 min read

Source: Claude Advisor Strategy Blog 2026 04 09 (Anthropic blog, Apr 9 2026 — https://claude.com/blog/the-advisor-strategy)

A new server-side tool on the Claude Platform (advisor_20260301, beta) that lets Sonnet or Haiku consult Opus mid-request when they hit a decision they cannot confidently resolve. Reverses the usual orchestrator-decomposes-to-workers pattern: the small, cheap executor drives, and escalates to the large model only when needed. Single /v1/messages request, no orchestration code.

Key Takeaways

  • Inverted delegation. Normal sub-agent setups have the big model orchestrate and spawn small workers. Here the small model is the driver and calls up to Opus as a consultant. Opus never calls tools or writes user-facing output — only returns a plan, correction, or stop signal.
  • Single API call. Declare advisor_20260301 as a tool in your Messages API request and handoff happens inside one /v1/messages round-trip. No context juggling, no extra infra.
  • max_uses cost control. Cap how many times per request the executor can consult the advisor. Advisor tokens report separately in the usage block.
  • Pricing model. Advisor tokens bill at advisor-model rates, executor tokens at executor rates. Since advisor plans are short (typically 400–700 text tokens) while the executor handles the full output, total cost stays well below running the advisor end-to-end.
  • Benchmarks (Sonnet executor + Opus advisor): +2.7 pp on SWE-bench Multilingual vs Sonnet solo, with 11.9% lower cost per agentic task. Also improved on BrowseComp and Terminal-Bench 2.0.
  • Benchmarks (Haiku executor + Opus advisor): 41.2% on BrowseComp vs 19.7% for Haiku solo — more than 2× Haiku’s solo score. Still 29% below Sonnet solo, but 85% cheaper per task, so it’s the pattern for high-volume workloads.
  • Beta header required. anthropic-beta: advisor-tool-2026-03-01.
  • Executor decides when to escalate. The small model chooses when to invoke the advisor tool — not a fixed policy from outside.

Implementation

  • Tool/Service: Claude Platform Messages API — advisor tool (beta)
  • Setup: Add anthropic-beta: advisor-tool-2026-03-01 header, add advisor_20260301 to your tools array, keep your executor model (Sonnet/Haiku) as the top-level model, set max_uses for cost ceiling.
  • Cost: Pay-as-you-go. Advisor tokens priced at Opus rates (~400–700 tokens per consult); executor at Sonnet/Haiku rates for everything else. Net cost lands near executor-only, per Anthropic’s own evals.
  • Integration notes: The advisor runs server-side — no client-side orchestration needed. Works alongside existing tools (web search, code execution, custom tools) in the same agent loop.
  • Code pattern:
response = client.messages.create(
    model="claude-sonnet-4-6",  # executor
    tools=[
        {
            "type": "advisor_20260301",
            "name": "advisor",
            "model": "claude-opus-4-6",
            "max_uses": 3,
        },
        # ... your other tools
    ],
    messages=[...]
)
# Advisor tokens reported separately in the usage block.

Why it matters

  • Third paradigm for multi-model agents. Managed Agents host the whole loop server-side. Agent Teams coordinate peer-level instances. Subagents decompose work downward to isolated workers. The advisor strategy is a fourth: upward consultation without decomposition.
  • Cost envelope is favorable. Sonnet solo beats Sonnet+advisor on raw throughput but the advisor config wins on both cost and quality for agentic tasks. That is unusual — normally you trade one for the other.
  • Right default for high-volume automation. Haiku + Opus advisor at 85% less cost than Sonnet solo is the sweet spot for something like Hermes or SEO content pipelines where many cheap decisions and occasional hard ones run through the same loop.

Customer validation (from announcement)

  • Bolt (Eric Simmons, CEO): “makes better architectural decisions on complex tasks while adding no overhead on simple ones”
  • Genspark (Kay Zhu, CTO): “better than a planning tool we built ourselves”
  • Eve Legal (Anuraj Pandey, ML Engineer): “matching frontier-model quality at 5× lower cost”

Open Questions

  • Advisor visibility to the executor’s tool results. Article says “Opus accesses shared context” — exact mechanics of what context is forwarded to the advisor on each consult aren’t specified (curated subset vs. full transcript).
  • Interaction with Managed Agents. Can Managed Agents declare advisor_20260301, or is this only for raw Messages API usage?
  • max_uses behavior at the cap. Does the executor get a soft signal (“advisor_budget_exhausted”) or just silently lose access to the tool?
  • Streaming support. Does advisor invocation work with streamed responses or require buffered completion?

Try It

  1. Smallest evaluation. Run your existing Sonnet agent on 20 representative tasks three ways: Sonnet solo, Sonnet + Opus advisor (max_uses=3), Opus solo. Compare quality, tokens, and cost. Anthropic recommends this three-way eval as the decision framework.
  2. High-volume switch. If you run a Haiku-based workflow at scale (tagging, routing, extraction), swap in Haiku + Opus advisor with max_uses=1 or 2. The BrowseComp numbers suggest double the quality for almost the same money.
  3. Ship gate for Hermes and SEO pipeline. The advisor pattern fits Hermes and SEO content pipeline workflows where most decisions are easy but a few are architectural. Candidate integration: add advisor to the skill-execution loop and measure whether “stuck” cases now resolve without human intervention.
  4. Read the official advisor docs at https://platform.claude.com/docs/en/agents-and-tools/tool-use/advisor-tool — Open Questions above are best answered there.

Graph View

Backlinks