一种智能体操作系统：每个应用都是一个 Markdown 文件
String: An Agentic OS Where Every App Is a Markdown File
浏览论文内容
中文总结 AI 辅助
该研究提出开源智能体操作系统 String，将应用定义为 SFMD 文档，通过分阶段视图和权限机制提升 LLM 智能体操作效率，在 87 任务基准测试中减少 token 消耗并保持良好成功率。
中文摘要 AI 辅助
大语言模型（LLM）智能体已成为一类新型软件用户，但它们所操作的所有界面都是为其他用户设计的。网页是为人类视觉设计的，人类可以快速浏览并忽略无关内容；工具模式是为程序设计的，程序无需为从未调用的定义付出任何代价。而智能体没有这种便利：它每次都要重新读取所显示的所有内容，并且每次交互都要为这些内容再次付费。我们提出了 String，这是一个开源运行时，它为这类新型用户提供了专属界面，并将该问题作为操作系统问题来处理。工具知识不再存储在智能体的上下文窗口中，而是移至公共层，该公共层每次以 Markdown 格式渲染回一个视图。单个 String 专用 Markdown（SFMD）文档声明应用的视图、类型化操作、导航和凭证，运行时通过两个核心动词处理发现、验证、执行、状态和密钥管理：/open 用于查看，/act 用于执行。结果表明，网页和应用是同一架构的两种渲染形式：SFMD 站点向浏览器提供样式化 HTML，向智能体提供原始文档，因此同一语法可覆盖应用、文件、 shell 和网页，甚至包括传统 HTML，无需为每个站点进行单独集成。视图设计为部分显示，且分阶段展示具有因果关系：提前一个交互阶段披露细节会导致准确率下降多达 23 个百分点，而适当的分阶段展示可将错误操作选择率从 28% 降至 2%。权限遵循来源：远程页面可调用 HTTP，但绝不能调用 shell；调用方提供的文本绝不能扩展存储的密钥。在包含 87 个任务的基准测试中，每个任务都配有精心设计的技能，将这些程序作为按需 String 应用实现后，在从前沿模型到小型模型的六种模型中，整体成功率相当（提升了 1.3 个百分点），同时在完成的交互轮次中使用的 token 减少了 33.5%，且常驻界面在任何目录大小下都保持恒定的 53 个 token。我们报告该设计、评估结果以及三个月生产使用的经验教训。
英文摘要
LLM agents have become a new class of software user, but every surface they work through was designed for someone else. Pages are built for human eyes, which can skim and ignore; tool schemas for programs, which pay nothing to carry definitions they never call. An agent has neither luxury: it re-reads, and pays again for, everything it is shown on every turn. We present String, an open-source runtime that gives this user an interface of its own and treats the job as an operating-systems problem. Tool knowledge moves out of the agent's context and into a common layer that renders it back one view at a time as Markdown. A single SFMD (String-Flavored Markdown) document declares an application's views, typed actions, navigation, and credentials, and the runtime handles discovery, validation, execution, state, and secrets behind two core verbs: /open to see and /act to do. Web and app turn out to be two renderings of one architecture: an SFMD site serves styled HTML to browsers and the raw document to agents, so one grammar reaches apps, files, shells, and the web, even legacy HTML, with no per-site integration. Views stay partial by design, and the staging is causal: disclosing one tier of detail a single turn too early costs up to 23 accuracy points, while proper staging drops wrong-action selection from 28% to 2%. Privilege follows provenance: a remote page may call HTTP but never the shell, and caller-supplied text never expands a stored secret. On an 87-task benchmark that pairs each task with curated skills, operationalizing those procedures as on-demand String apps yields comparable aggregate success across six models from frontier to small (+1.3pp) while using 33.5% fewer tokens among completed episodes, and the resident interface stays a constant 53 tokens at any catalog size. We report the design, the evaluation, and what three months of production use taught us.
发表机构
- Seoul National University（首尔大学）
- H1R.AI
机构由 AI 辅助整理，请以论文原文为准。
