Research projects & roadmap研究项目与路线图
ClawBench Series
Real tasks. Working software. Learning agents.
真实任务，可运行的软件，持续学习的 Agent。
A series of projects on how agents understand, operate, reconstruct, and learn from digital environments—from live websites to desktop software.
围绕 Agent 如何理解、操作、重建数字环境，并从中学习展开的一组项目， 从真实网站延伸到桌面软件。
The common thread共同主线
From evaluation to environments.从评测出发，走向环境与学习。
ClawBench starts with agents doing useful work on the live web. WebsiteBench asks how those websites can be reconstructed, while ClawBenchPro brings task execution into reproducible web environments.
REABench extends the question to understanding software behavior. REAGym explores turning software into training environments, and CUABench focuses on the work that computer-use agents can complete.
ClawBench 从 Agent 在真实网站上完成实际任务出发；WebsiteBench 研究如何重建网站， ClawBenchPro 则把任务执行放进可复现的网站环境。
REABench 将问题延伸到软件行为理解；REAGym 探索把软件变成训练环境， CUABench 关注电脑使用 Agent 能够完成怎样的工作。
The projects系列项目
Public projects & work in progress公开项目与正在推进的方向ClawBench
Public project公开项目Can agents complete everyday tasks on real websites?
Agent 能否在真实网站上完成日常任务？
ClawBench evaluates browser agents on live websites. Tasks involve real workflows such as shopping, scheduling, and applications; final submissions are intercepted for evaluation before they take effect.
ClawBench 在真实网站上评估浏览器 Agent，任务覆盖购物、预约、申请等实际流程， 在最终提交生效前拦截请求并评估任务是否完成。
- Setting环境
- Live websites真实在线网站
- Focus研究重点
- Task completion in real-world web workflows真实网页工作流中的任务完成能力
WebsiteBench
Public project page项目主页已公开Can agents rebuild a website through browser exploration?
Agent 能否通过浏览器探索重建一个网站？
WebsiteBench studies website reconstruction from browser interaction. An agent explores a reference website and builds a working implementation, with evaluation focused on functionality as well as appearance.
WebsiteBench 研究通过浏览器交互重建网站：Agent 探索参考网站， 实现可运行的版本，评估同时关注功能行为与视觉呈现。
- Setting环境
- Browser exploration and website implementation浏览器探索与网站实现
- Focus研究重点
- Recovering how a website looks and works还原网站的界面与功能
ClawBenchPro
In development开发中Can agents complete workflows in reproducible web environments?
Agent 能否在可复现的网站环境中完成工作流？
ClawBenchPro develops task evaluation on locally hosted websites. Controlled environments make it possible to repeat workflows and check the resulting application state.
ClawBenchPro 面向本地托管网站构建任务评测， 通过受控环境重复执行工作流，并检查任务完成后的应用状态。
- Setting环境
- Self-hosted website environments自托管网站环境
- Focus研究重点
- Repeatable workflows and verifiable outcomes可重复的工作流与可核验的结果
REABench
In development开发中Can agents recover and reproduce software behavior?
Agent 能否理解并复现软件行为？
REABench focuses on reverse-engineering tasks grounded in real software. Agents investigate a software artifact and reproduce selected behavior in executable code.
REABench 关注真实软件上的逆向工程任务， 让 Agent 分析软件工件，并用可执行代码复现选定的功能行为。
- Setting环境
- Software artifacts and executable behavior软件工件与可执行行为
- Focus研究重点
- Behavioral understanding and functional reproduction行为理解与功能复现
REAGym
In development开发中How can software become a reinforcement-learning environment?
如何把软件变成强化学习环境？
REAGym develops interactive training environments from software and games. The goal is to expose observations, actions, rewards, and resettable episodes so that agents can learn through interaction.
REAGym 从软件和游戏中构建交互式训练环境， 提取观测、动作、奖励与可重置的回合，让 Agent 通过交互学习。
- Setting环境
- Interactive software and game environments可交互的软件与游戏环境
- Focus研究重点
- Environment construction and agent learning环境构建与 Agent 学习
CUABench
Research direction研究方向Can computer-use agents finish multi-step work across applications?
电脑使用 Agent 能否跨应用完成多步骤工作？
CUABench focuses on evaluating computer-use agents in application workflows: navigating interfaces, managing application state, and producing the work that a task calls for.
CUABench 面向应用工作流评估电脑使用 Agent， 关注界面操作、应用状态管理，以及最终工作成果的完成。
- Setting环境
- Computer-use workflows电脑使用工作流
- Focus研究重点
- Work completion across application interfaces应用界面中的实际工作完成能力