Strategist evolves reusable inference strategies at 39% lower cost — Inference strategies as executable code — What does it mean?

The news. On November 2, 2026, researchers at Princeton University and Meta posted Strategist, a meta-agent that evolves inference strategies across tasks. On 30 benchmarks it beat the strongest of 18 baselines on 24, reaching 74.7% average accuracy against 74.0% at 39% lower inference cost. Read the paper →

Picture a kitchen where every new dish is invented from scratch and the recipe is thrown away after service. That is how most agent systems are built: a person writes the inference strategy once per system, and nothing learned carries to the next one. Strategist keeps the recipe binder instead. Each strategy is saved as a blueprint: a Python program that fixes the structure (for example, draft three answers, review each, take a majority vote) and leaves the prompts out. Fill in prompts and the blueprint becomes a runnable strategy.

Because every blueprint shares the signature execute(llm, task), any blueprint can call any other, so a new recipe can use an old one as a sub-recipe. Two foundations sit at the bottom, chat (single or multi-turn conversation) and agent (a scaffold with tools). Four rules keep the binder tidy: one shared interface, no calls to blueprints written later, no cycles, and no duplicates.

The head chef here is a meta-agent, and it runs four steps per round with frozen models and no training. Retrieve: search the library by measured behavior and past notes (TF-IDF plus embedding similarity), and also pull in adjacent blueprints from other domains as inspiration. Propose: write candidates that are new, composed from an existing blueprint, or a reuse with new prompts; every round must include at least one of each. Evaluate: run them on the training split with full traces, the validation split with scores only, and the hidden split to pick a winner; a fixed test split stays untouched. Accumulate: store the blueprint, structure only, with its measurement card and notes.

Dropping the prompts when storing is what lets a blueprint move to a task it was never built for. The search loop runs 10 iterations of five candidates each. For the same reason, the meta-agent can read why a card scored well before it borrows it, much as a cook reads the tasting notes before reusing a sauce. See when more reasoning tokens are worth it for the budget question each blueprint is answering.

The other half of the saving is routing inside one strategy. Every blueprint step can target a cheaper or costlier tier, so a strategy can be accurate and cheap at the same time, by spending the head chef only where it changes the answer. The paper's example for an Amazon recommendation task builds a user profile with cheap calls, then sends only users whose preferences are shifting to an expensive multi-agent debate. The branch itself is plain deterministic code that decides which inputs deserve the expensive call, and because the result is Python, a person can read and edit it. This is the lever behind the cost profile of an agent.

Here is how routing composes into a saving, with every number illustrative. Hold three things fixed: 200 test questions, a cheap call costing 1 unit, and an expensive call costing 10 units. A uniform strategy sends every question to one expensive call: 200 × 10 = 2,000 units. A routed strategy makes one cheap triage call per question (200 × 1 = 200 units) and escalates 20% of them, 40 questions, to the expensive call (40 × 10 = 400 units). The total is 600 units, a 70% cut, but only if the triage keeps the 160 skipped questions as accurate as before. That condition is exactly what the evaluate step measures against the hidden split. The paper's measured average over 30 benchmarks is a 39% lower weighted cost at matched accuracy, so real strategies save less than this best-case sketch.

Read the result with its limits. Strategist wins on 24 of 30 benchmarks, ties on 2, trades accuracy for cost on 2, and loses on 2 (HotpotQA and COPA, at equal accuracy but 9% and 6% higher cost). The winning strategy is chosen for being best on the hidden split, so part of its margin there is luck; the test split, where nothing is chosen, is the only unbiased estimate. On 10 held-out benchmarks, retrieving a blueprint from the library with no search stayed within 2 points of the strongest baseline on all 10 and beat it on 8, and strategies found on Qwen3 still beat the baselines when run unchanged on GPT-5 and Gemini-2.5.

Goes deeper in: AI Agents → Planning & Reflection → When to Spend More Tokens