Strategist: a meta-agent that evolves reusable reasoning strategies — 24/30 benchmark wins at 39% lower inference cost
A meta-agent writes reasoning strategies as executable blueprints, accumulates them in a library, and transfers them to unseen tasks and models without re-searching.
Princeton and Meta researchers have built a meta-agent that writes the reasoning strategies other agents use, stores them as reusable executable blueprints, and lets later tasks — and even other models — reuse them without re-searching. On 30 benchmarks, Strategist beats the strongest baseline on 24 and matches its average accuracy (74.7% vs 74.0%) at 39% lower inference cost, per the paper (arXiv:2611.03618).
What shipped
Strategist is a meta-agent: it does not answer questions; it designs the inference procedure that does. Each strategy is a Python "blueprint" with a single shared interface, execute(llm, task) -> str, built from two foundations — chat (single- or multi-turn conversation) and agent (scaffold plus tools). A blueprint captures reasoning structure — steps, control flow, which model tier runs each step — but not task-specific prompts; filling in the prompts yields a runnable strategy (full text).
The library enforces four protocols: a shared interface, no forward calls (a node can only call nodes created before it), no cycles, and no duplicates. Strategies route across three model tiers (fast, balanced, complex), each in two modes (nothink, thinking). In the released code those tiers are Qwen3-0.6B, Qwen3-1.7B, and Qwen3-4B (repo).
The loop runs four stages per iteration:
- Retrieve relevant blueprints plus "adjacent inspiration" via hybrid TF–IDF + embedding search.
- Propose K=5 candidates as a mix of
new,composed, andreuseblueprints (at least one of each per iteration). - Evaluate on real samples — full traces on the training split, only aggregated scores on validation.
- Accumulate the underlying blueprints (prompts discarded) into the library with measurement cards and experience notes.
The data is split four ways: training (full traces seen), validation (aggregated scores only), hidden (selects the winning strategy across iterations), and test (held out entirely, final confirmation only).
Why it matters
The key claim for people building agent systems is the cumulative one: strategies are not optimized per-task and thrown away, but stored as composable, prompt-free blueprints that later tasks reuse. The paper reports two concrete transfer results (full text):
- On 10 held-out benchmarks, one-shot library retrieval without any search stays within 2 accuracy points of the strongest baseline on all 10 and beats it on 8 (mean 77.8% → 81.4%).
- The same strategies, moved to GPT-5 and Gemini-2.5 without changing a line, still beat the baselines there.
The reasoning procedure — normally hand-written and never reused — becomes a searchable, accumulating asset. The library ablation backs the mechanism: with the library, 10-benchmark test accuracy is 78.7%; ten-iteration search with the library disabled drops to 65.4%, a 13.3-point gap (alphaXiv).
The cost numbers are also worth reading carefully. The "39% lower" figure is a mean per-benchmark weighted inference-token cost of 0.61× the baseline, where the fast/balanced/complex tiers are weighted 1×, 2.83×, and 6.67×. Example: on 2WikiMQA, Strategist hits 76.5% vs 72.5% at a weighted cost of 33.8k vs 98.6k tokens per sample (0.34×) (alphaXiv).
Caveats
These are the authors' own reported results, and no independent evaluation turned up in the sources checked — so treat the headline numbers as vendor claims, not established fact.
The sharpest caveat is what "39% cheaper" does not mean. It is per-sample inference cost after search, not elapsed time and not the cost of finding the strategy. The search loop itself is computationally substantial — it runs a meta-agent over a development set across iterations. The honest engineering question is whether the accumulated library amortizes that up-front search cost across enough tasks to pay for itself.
The paper's stated limitations, as summarized by alphaXiv, concede that the underlying models never learn, gains depend on development data and a search budget, and the benefit is an average over benchmarks, not a guarantee on each one.
Two more cautions. First, the selection-vs-test gap: on the hidden selection split the chosen strategy sits 4.3 points above baseline, but on the held-out test only 0.7 points above — the test figure is the unbiased estimate. Second, the library ablation removes retrieval, reusable components, and accumulated notes together, so it does not isolate "inspiration" as the cause, and it runs on 10 benchmarks, not the 30 in the headline.
Also unconfirmed from my sources: which strategist LLM produced the headline results (the repo lists Codex CLI, Claude Opus, and a Claude Code CLI on MiniMax as options), the full 30-benchmark list, and the 18 baseline names.
