Tag
作者探讨为什么移动智能体（mobile agents）的表现远不如编码智能体（coding agents），指出手机具备摄像头、传感器、位置、通知和大量个人上下文等丰富资源，但现有移动智能体多为普通AI助手套壳或仅点击屏幕的简单实现，并提出关于限制因素（模型能力、系统限制、工具链或问题定义方式）的开放讨论。
The paper identifies why MCTS with critics fails to improve computer-use agents and proposes Probe-and-Resume, a policy that invokes a critic only on candidate final states with full probe access to a forked environment, boosting Terminal-Bench 2.1 success by 8.4 points at less than half the cost of Codex CLI.
Tencent ARC 的论文提出 Player-1 一套面向游戏控制的后训练框架，通过扩展动作空间、在 6,044 条游戏轨迹上进行 SFT 初始化，并利用 16 个可程序化验证的合成环境进行 PPO 多环境强化学习，显著超越监督基线与现有最强 baseline。
CLASHarness introduces a controllable, verifiable framework for generating multi-agent conflict scenarios in computer-use tasks, benchmarking 15 frontier vision-language models and revealing weaknesses in ground-truth verification, conflict resolution, and shared-resource coordination.
用户 rileybrown 称赞 Codex 的 computer use 功能非常强大，表示现在已经离不开它了。
Poorna Singh Baghel 介绍 Betrox AI：一个多 worker 架构的自主智能体系统，可将单条高级自然语言指令拆解并交给专门的后台 worker，同时操作桌面应用、浏览器、终端和本地文件，实现完整的操作系统级控制。
A tweet highlights a benchmark showing a big gap in computer-use ability between large general models and specialized models, which mostly fail at pointing tasks due to lacking the necessary vocabulary.
A brief X post joking that computer use appears to be "solved," quoting a comparison showing that large general models point on computer UIs better than specialized models due to richer vocabulary.
作者重新评估Jev类模型的价值，认为随着模型足够快和便宜，原本靠if-else拼接的业务逻辑、实时语音交互、计算机操作和机器人决策都可以交给System 1模型处理，形成SOTA后台深思考模型与System 1实时模型分工的新范式。
Pine AI 发布 Pine Computer，宣称通过将 computer use 视为编码问题并重构智能体观察空间，仅需 $1 成本即可达到竞品 $25 的效果，且启动速度快 2-5 倍，有望解决 computer use 太贵太慢的痛点。
ViSkill 提出一种视觉原生的技能学习框架，将成功交互编码为可复用的视觉技能卡，与策略优化形成闭环反馈，在 Sokoban、FrozenLake 等基准上超越多种基线。
Opposable 发布于 Product Hunt，让 Claude Code、Codex 或任意 MCP 智能体能够看到、点击并输入真实 iPhone 或 Android 手机上的内容，覆盖所有已登录应用；常见任务在手机端直接运行无需模型，支付需用户确认，iPhone 可通过蓝牙或 USB 小配件连接。
offstage 是一款免费开源工具，为 Claude Code、Codex、opencode 等编程智能体提供一个独立的 macOS 登录账户，让模拟器、Xcode UI 测试和智能体构建的应用在该账户的桌面上运行，从而不占用用户自己的屏幕、键盘和鼠标。
Figure founder Brett Adcock announced the launch of HARK Pro, a personal AI agent built on Handoff, a computer-use model that lives in the cloud. It can autonomously complete real-world tasks such as shopping, calling a ride, ordering food, and compiling data. Available today on web, iOS, and Android, with subscriptions priced at $20/month or $200/year.
Startup Hark, founded by serial founder Brett Adcock, has publicly released Hark Pro, an AI personal assistant built on a computer-use model that operates users' computers end-to-end while emphasizing privacy over ad-driven data monetization. The product features a custom interface with chat, proactive task feeds, and a visible browser-navigation window designed to build user trust.
Gumloop launches Agent Browsers, a hosted browser environment that lets AI agents operate on sites lacking an MCP or API — with cloud-run browsers, credential vaults with 1Password integration, persistent profiles, live viewing and session replays, stealth/proxied sessions, and reusable workflows.
OpenAI announced a partnership with Ironclad to turn real contracting and procurement workflows into training and evaluation tasks for AI agents, with GPT-6 Astra scoring 32% higher than GPT-5.6 Sol on these tasks while cutting estimated time per attempt by 48%.
作者描述 Windows 10 因被视为已停止支持而遭 OpenAI 和 Anthropic 放弃兼容性更新，导致 ChatGPT 的 Computer Use 无法截屏、Work 模式报错，以及 Claude 的 shell 功能在 W10 更新后失效，逼迫用户升级到不愿使用的 Windows 11。
宝玉 (@dotey) 深入分析了 amontlabs/lcu 项目：它并非自研模仿，而是直接复用 OpenAI ChatGPT 桌面版内置的 Computer Use 运行时，通过 MCP 将该能力暴露给 Claude Code、Codex CLI、Pi 等任意 Agent 工具调用。LCU 仅提供 js 和 js_reset 两个工具，模型通过编写 JavaScript 操作预置的 cua 对象（getState/getApp/click/typeText 等）完成桌面控制，并支持访问性 API 直接点击 UI 元素。
Reduck 推出 Reduck MCP，作为浏览器操作的记忆层，通过复用脚本让 AI 代理的浏览器任务提速约 2.5-3 倍，并借助 Chrome 扩展解决登录与反机器人检测问题，同时维护主流网站的官方脚本库。