Strategist: Evolving Reasoning with Agents
Taiming Lu and Zhuang Liu (Princeton University) with Qi Xu and Zhigang Hua (Meta) introduce Strategist, a meta-agent that writes inference strategies as executable code, tests them on real samples, and keeps every strategy in a shared library that later tasks reuse.
Ask this paper
Problem. The inference strategy, the reasoning procedure that turns model calls into a response, is usually written by hand, and what was learned building one strategy does not carry over to the next task.
Blueprints. Each strategy is a Strategy Blueprint, an executable program with one shared interface, so any blueprint can be used as a component inside another. Blueprints and their measured scores form a library that grows as it is used.
Search loop. Given a task and a small development set, Strategist retrieves related blueprints, proposes five candidates, evaluates them on real samples and revises them over ten iterations. The strategist model (GPT-6 via the Codex CLI) only writes strategies; frozen task models answer the questions, with no training.
Results. Against 18 baselines from direct prompting to multi-agent workflows, Strategist beats the strongest baseline on 24 of 30 benchmarks and matches its average accuracy (74.7% vs 74.0%) at 39% lower inference cost. One-shot retrieval from the library alone reaches 81.4% vs 77.8% at 23% lower cost.
Transfer. Found strategies transfer to other executors without new search, adding 1.6 points on Qwen3, 20.5 on GPT-5 and 14.7 on Gemini-2.5 at 0.60x to 0.65x the cost. The train-to-hidden gap stays between 1.2 and 2.5 points across iterations.
Abstract
A model is only half of an AI agent system. The other half is the inference strategy: the reasoning procedure that turns model calls into a response. Today that second half is written by hand, and the experience of building one strategy never carries to the next. We introduce Strategist, a meta-agent that evolves inference strategies across tasks. Each strategy is written as a blueprint,executable code that other strategies can reuse, and together the blueprints form a library. Given a task and a small development set, Strategist retrieves related blueprints, writes candidate strategies, tests them on real samples, and revises them from the results. Every candidate then enters the library, so later tasks build on earlier discoveries. 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. Its strategies also transfer to unseen tasks and other models without further search. The part of an AI system that is not improved on its own now does.