编码代理设计框架实证研究发布

一项针对编码代理设计框架的实证研究于2026年9月17日发布1，该研究通过在四个大语言模型上评估176个配置，对自主编程系统的核心设计要素进行了系统分析1。研究在SWE-Bench Verified和Terminal-Bench 2.1两个基准上进行评估1，重点考察了规划、动作空间和上下文管理三个组件对编码代理性能的影响。

研究发现上下文管理在上下文窗口预算紧张时具有显著价值，其主要作用在于防止溢出失败1。在多种上下文管理策略对比中，规则式删除与大语言模型摘要相结合的混合方法表现最优1。此外，规划机制对不同能力等级的模型影响存在差异——对于强模型，规划主要作用是降低成本而非提升精度1；而在动作空间选择上，bash能力较强的模型采用bash接口可实现最低的接口成本1。

Researchers have conducted an empirical study evaluating how different design frameworks affect the performance of autonomous coding agents.1 The investigation assessed 176 distinct configurations across four large language models, focusing on three core components: planning, action space, and context management.1 The evaluation used SWE-Bench Verified and Terminal-Bench 2.1 as benchmarks.1

The study was submitted on September 17, 2026.1 Key findings reveal that context management strategies deliver significant value when operating under tight context window budgets, primarily by preventing overflow failures.1 Among the context management strategies tested across four different budget levels, a hybrid approach combining rule-based deletion with language model summarization proved most efficient.1

The research also found that planning affects models differently depending on their capabilities: for stronger models, planning primarily reduces computational costs rather than improving accuracy.1 Additionally, models with robust bash capabilities achieve the lowest interface costs when using bash-based interaction.1