a builder's codex
codex · operators · Cat Wu · ins_advisor-tool-replaces-ensemble

Advisor-tool replaces ensemble-of-3 stability hacks at near-Sonnet rates

By Cat Wu · Head of Product, Claude Code + Co-work, Anthropic · 2026-04-26 · essay · Anthropic Advisor Tool — public beta documentation

Tier B · TL;DR
Advisor-tool replaces ensemble-of-3 stability hacks at near-Sonnet rates

Claim

Anthropic's advisor-tool collapses the ensemble-of-3 stability pattern into a single executor run that consults Opus mid-stream for a 400-700 token plan, then continues at Sonnet rates. Valid pairs: Haiku/Sonnet/Opus 4.6 → Opus 4.7 advisor. For long-horizon, mostly-mechanical workloads with occasional planning bottlenecks, this is the cleaner replacement for ensemble voting.

Mechanism

Ensembles deliver stability by running multiple full inferences; the advisor pattern delivers it by inserting a small Opus-grade plan at the moment of decision, paid for once instead of N times. Cost-per-stable-decision improves because the executor stays on a cheap tier for the bulk of the work.

Conditions

Holds when: the workload has clear planning bottlenecks (consolidation, multi-step decisions) that benefit from a higher-capability mid-stream consult.

Fails when: every step needs high-capability reasoning — there, ensemble or pure Opus still wins.

Evidence

Anthropic advisor-tool public beta with header advisor-tool-2026-03-01. Sonnet executor calls Opus mid-generation for a 400-700 token plan. Verified via Anthropic Python SDK 0.97.0.

— Anthropic platform docs, 2026-04-26

Signals

Counter-evidence

For workloads requiring diversity-of-output (where ensembles vote across genuinely independent reasoning paths), advisor cannot substitute — it is one path with a smarter midpoint.

Cross-references

Open the interactive view → View original source → Markdown source →