# Strategist: Evolving Reasoning with Agents

Source: [arXiv](https://arxiv.org/abs/2611.03618v1)  
Feed7 permalink: https://feed7.dev/p/2611-03618v1-02clswt  
Published: 2026-11-02T18:51:04.000Z  
Trust: Needs Review (needs_review)

## Why Included

Strategist uses a reusable library of executable reasoning blueprints that agents retrieve, test, and revise across tasks. The paper reports similar average accuracy to its strongest baseline at lower inference cost.

## Source Summary

**Strategist** retrieves related executable blueprints, writes candidate inference strategies, tests them on a small development set, and revises them. New candidates join a shared library for later tasks. Across **30 benchmarks**, it reports 74.7% average accuracy versus 74.0% for the strongest baseline, at 39% lower inference cost.

## Practical Implication

For multi-task agent systems, consider storing reusable strategy code and testing it against task samples before deployment. The authors also report transfer to unseen tasks and other models without additional search.

## Agent-Ready Context

**Strategist** retrieves related executable blueprints, writes candidate inference strategies, tests them on a small development set, and revises them. New candidates join a shared library for later tasks. Across **30 benchmarks**, it reports 74.7% average accuracy versus 74.0% for the strongest baseline, at 39% lower inference cost.

For multi-task agent systems, consider storing reusable strategy code and testing it against task samples before deployment. The authors also report transfer to unseen tasks and other models without additional search.

These are paper-reported benchmark results; the supplied material does not describe the benchmark mix or implementation details needed to judge fit for a particular coding-agent workflow.

## Connected Context

Feed7 judgment across 990 accumulated Signals:

This describes a reusable strategy library built from executable blueprints that are tested on development samples and revised, with reported benchmark accuracy and inference-cost results. SkillProx adds a maintenance implication: measured retries, rollback, and pruning can help govern accumulated reusable procedures. The supplied material does not establish that either approach fits coding-agent workflows.

- [SkillProx: Self-Evolving Agent Skills via Proximal Textual Gradient Descent](https://feed7.dev/p/2608-07449v1-1co8km3) — Both evolve reusable agent guidance through evaluation. Strategist tests executable strategies on development samples and adds candidates to a shared library; SkillProx highlights regression rollback and utility-based pruning as safeguards for maintaining such accumulated knowledge.
- [GEPA-Curator: Curated Learning for Self Improving Agents](https://feed7.dev/p/2610-37651v1-1usz0jw) — GEPA-Curator retrieves successful trajectories before prompt optimization, while Strategist retrieves and tests executable blueprints; both make prior successful work reusable, though they apply it through different mechanisms.

## Context Map

- Layer: agent
- Domains: coding, research
- Topics: harness-engineering, skills, agent-evals

## Uncertainty

- These are paper-reported benchmark results; the supplied material does not describe the benchmark mix or implementation details needed to judge fit for a particular coding-agent workflow.

## Agent Instruction

Use this item as source-backed context. Do not invent claims beyond the linked source. If this item conflicts with another source, call out the conflict.