Exo:面向系统运行时持续自我进化的智能体架构
Exo: An Architecture for Self-Improving Systems at Runtime
arXiv: 2611.00463 [cs.AI]
Submitted: October 30, 2026
Authors: Alexander Krentsel, Sylvia Ratnasamy, Ion Stoica, Martin Casado, Ankur Goyal
文章背景与核心概要
随着大语言模型 (Large Language Model, LLM) 在实际系统中的广泛部署,AI 智能体 (AI Agent) 的环境治理机制 (Harness,即规范模型如何与外部环境安全交互的控制中枢) 成为决定任务成败与运行成本的关键。然而,传统的手工调优极其繁琐昂贵,现有的自动化优化手段又强依赖静态数据集和高开销的离线执行。来自 UC Berkeley 等机构的研究团队提出了创新的运行时智能体架构 Exo,其核心突破在于将优化循环直接折叠进正在执行的任务流程中,使智能体能够在实际工作中安全、自主地持续修改自身的治理机制与求解策略。通过极简受保护机制层对执行历史与策略激活进行安全隔离与审计,Exo 在开放世界游戏、AppWorld 以及 SREGym 等多个极具挑战性的基准测试中,将模型成本削减高达 12% 至 82%,大幅缩短了任务排障与执行时间。这一工作标志着 AI 智能体正式由“离线预设固定系统”迈向“线上自主协同进化”的新阶段。
📌 内容概要
📌 Summary
对于 AI 智能体而言,纯手工编写环境治理机制 (Agent Harness,用以规范大语言模型与外部环境交互行为的控制策略) 是一项极其复杂的系统工程;而传统的自动化优化手段,往往需要专门构建代表性的离线数据集并承担高昂的重复执行成本。本文提出了 Exo——一种专为智能体在日常实际执行过程中实现安全学习与自适应演化而设计的创新运行时架构。
Designing agent harnesses (which govern how language models interact with environments) manually is exceptionally difficult, while traditional automated techniques demand representative datasets and costly executions. This paper introduces Exo, an innovative agent runtime architecture designed for safe learning during ongoing work.
通过将优化反馈循环直接嵌入并压缩到任务执行流程之中,Exo 赋能智能体在运转中动态迭代其环境治理机制以及解决问题的具体策略。系统内部设计了一个极简的受保护机制层,在安全封存完整执行历史的同时严密把控策略的激活生效,确保环境治理机制通过完全透明、可审查的代码变更实现平稳进化。
By collapsing the optimization loop directly into task execution, Exo empowers agents to dynamically revise both their harnesses and problem-solving strategies. A minimal protected mechanism layer safely preserves execution history and governs policy activation, ensuring that the harness evolves smoothly through inspectable code changes.
核心性能亮点
Key Performance Highlights:
- 开放世界游戏任务 (涵盖长周期复杂目标):
- 在独立运行测试中,相较于 Codex 与 Pi,将模型调用成本降低了 12–53%。
- 在迁移运行测试中,在保持同等任务完成率的前提下,模型调用成本显著降低了 45–82%。
- AppWorld 与 SREGym 基准测试:
- 与无法编辑策略的静态初始方案相比,在保持同等正确率的同时,将保留测试集上的模型成本降低了 35–61%。
- 将 AppWorld 的平均任务耗时缩短了 13%。
- 将 SREGym 的故障诊断耗时中位数降低了 35%。
- Continual Learning Bench 编程基准测试:
- 将总模型成本 (包含学习阶段的消耗) 降低了 14%,且两种策略均成功解决了全部测试问题。
- Open-World Game Tasks (Long-running objectives):
- Reduced model cost by 12–53% relative to Codex and Pi in standalone runs.
- Reduced model cost by 45–82% in transfer runs, while matching task completion rates.
- AppWorld & SREGym Benchmarks:
- Cut held-out model cost by 35–61% (with comparable correctness) compared to a static starting policy.
- Reduced AppWorld's mean task time by 13%.
- Reduced SREGym's median diagnosis time by 35%.
- Continual Learning Bench Coding Tasks:
- Lowered total model cost (including the learning phase) by 14%, with both policies successfully resolving every issue.
📝 论文摘要
📝 Abstract
智能体环境治理机制 (Agent Harness) 决定了大语言模型在具体环境中的行为方式;然而,依靠人工设计这些治理策略不仅难度极高,现有的自动化技术更依赖代表性的静态数据集以及昂贵的执行开销。在此,我们提出了 Exo——一种面向任务执行期间进行安全学习的智能体运行时架构。该架构将传统的优化循环直接折叠进任务执行中,允许智能体在推进工作的过程中动态修正其环境治理机制与后续演化策略。系统通过一个极简的受保护机制层保存执行历史并管控策略激活,确保治理机制能够以完全透明、可审查的代码变更形式实现稳步进化。在开放世界游戏的长周期目标测试中,相比于 Codex 和 Pi,Exo 在独立运行测试中将模型成本降低了 12% 至 53%,在迁移运行中降低了 45% 至 82%,同时保持了完全匹配的任务完成率。与不具备策略修改能力的相同初始策略相比,Exo 在 AppWorld 和 SREGym 基准测试中,在保持相近正确率的同时将保留测试集的模型成本降低了 35% 至 61%,使 AppWorld 的平均任务耗时减少了 13%,并将 SREGym 的中位数诊断时间缩短了 35%。在 Continual Learning Bench 编程基准测试中,即使将学习阶段的开销计入在内,模型总成本依然下降了 14%,且两种策略均成功解决了所有挑战。
Agent harnesses govern how language models act in an environment; designing their policies by hand is difficult, while automated techniques require representative datasets and costly executions. We present Exo, an agent runtime architecture for safe learning during ongoing work that collapses the optimization loop into task execution, allowing agents to revise both their harness and their strategies for further improvement. A minimal protected mechanism layer preserves execution history and governs policy activation while the harness evolves through inspectable code changes. On long-running objectives in an open-world game, Exo reduces model cost by 12-53% relative to Codex and Pi in standalone runs and 45-82% in transfer runs, while matching task completion rates. Compared with the same starting policy without the ability to edit policy, Exo reduces held-out model cost by 35-61% on AppWorld and SREGym with comparable correctness, cutting AppWorld's mean task time by 13% and SREGym's median diagnosis time by 35%. On Continual Learning Bench coding tasks, model cost including learning falls by 14%, with both policies resolving every issue.
🔗 更多资源与链接
🔗 Additional Resources
- 全文链接: 查看 PDF | HTML 网页版 (实验性) | TeX 源码
- 数字对象唯一标识符 (DOI): 10.48550/arXiv.2611.00463
- 授权协议: 知识共享署名 4.0 国际许可协议 (Creative Commons Attribution 4.0 International, CC BY 4.0) (授权图标见下方)

- Full-Text Links: View PDF | HTML (Experimental) | TeX Source
- Digital Object Identifier (DOI): 10.48550/arXiv.2611.00463
- License: Creative Commons Attribution 4.0 International (view license icon below)