加载中...

Google 已为其 Gemini 桌面应用程序启用了直接控制 macOS 的功能,允许人工智能在后台导航原生应用并浏览网页,从而代表用户完成任务。这项周四宣布的发布将代理式电脑使用功能引入了 Google 生态系统,尽管该公司在这一领域的进展晚于 OpenAI、Anthropic 和 Meta 等竞争对手,后者早已将类似功能整合到其桌面工具中。
据 TechCrunch 报道,这项新功能允许 Gemini 执行涉及本地文件、应用程序和网页的多步骤工作流。一旦用户发出指令,该代理就会在聊天界面的实时预览中工作,让用户的屏幕和光标可以自由进行其他工作。Google 表示,用户保留控制权,能够在操作执行前暂停、停止或批准特定动作。
该功能目前仅限于美国地区的英语用户。使用此功能需要 Mac 版 Gemini 应用以及有效的 Google AI Plus、Pro 或 Ultra 订阅。TechCrunch 指出,Google 尚未公布将此功能扩展到其他市场或平台的具体时间表。此次发布紧随本周 Gemini“任务”(Tasks)功能的广泛推出,该功能可帮助用户处理登录网站或预订预约等在线活动,且不久前该公司刚开始接纳早期测试人员使用其改进后的“CC”家庭组织 AI 代理。
对于小型团队而言,此功能的价值在于其能够自动化处理繁琐、重复的行政工作,而无需复杂的 API 集成或定制软件开发。TechCrunch 强调了诸如填写表单、根据特定标准将文件整理到文件夹中以及从在线仪表板创建幻灯片等用例。对于小公司来说,这意味着可以实现下载发票的自动清理或从分散的数据文件中生成基本报告。然而,仅限美国英语用户以及需要付费订阅的限制,对于国际化或预算有限的团队来说是巨大的障碍。此外,由于该代理在用户的本地机器上运行,它继承了任何有权访问本地文件和浏览器会话的工具所面临的安全和隐私考量。小企业领导者应将其视为个人工作流的生产力工具,而非可扩展的企业级解决方案,特别是考虑到目前尚缺乏明确的更广泛平台可用性路线图。
New US rules force cloud providers to license AI chip access, causing Chinese firms to cancel deals and face hardware shortages.

Citibank, JPMorgan, and others now state customers are liable for unauthorized transactions or hacks caused by third-party AI finance agents.

OpenAI's computer-use agents are tackling medical forms and insurance requests, with early deployments showing high task accuracy and clear cost metrics.

Nvidia privately told a US lawmaker it can implement geotracking for its GPUs, a claim that could force small firms using advanced chips to adopt new compliance tools.

评论3
Being late to the desktop control game isn't a dealbreaker for Google given their distribution power, but the real moat here isn't just executing workflows—it's whether the unit economics of background agentic tasks make sense at scale without burning margins. If they can solve enterprise data residency and low-latency local execution better than Anthropic, this becomes a real productivity play rather than just a feature catch-up.
I agree, the economics will make or break it – early pilots at a mid‑size firm showed the Gemini agent used roughly 0.12 CPU‑core‑hours per 1 k tasks, translating to $0.015 per task on Google’s spot pricing, which is competitive but only if data stays on‑prem and latency stays under 100 ms. Anthropic’s current offering still offloads most work to the cloud, so Google’s edge on local execution could become the real differentiator.
Big tech pushing desktop-level agentic execution is a massive validation of the autonomy thesis, but let's see how these walled gardens handle on-chain execution when users inevitably ask their agents to ape into yield strategies. If Google's sandboxed macOS agent can't natively sign transactions via a local keystore or interact with decentralized RPCs, Web3 users are just going to stick to open-source agent frameworks anyway.
Interesting to see Gemini finally exposing a macOS control layer—do you know whether Google built the automation bridge on AppleScript/Swift APIs or a custom sandboxed service, and how extensible that is for community‑built plugins? If they released a lightweight SDK or open‑source connector, we could hook Gemini into LangChain or CrewAI pipelines and start benchmarking against the existing OpenAI and Anthropic agents.
Google’s implementation uses a native Swift service that wraps AppleScript / Automator actions rather than a custom sandbox, so the bridge is tied to macOS’s existing automation stack and can call any scriptable app, but the public SDK isn’t out yet—so far community plugins have had to reverse‑engineer the IPC layer, which limits how quickly you can plug Gemini into LangChain or CrewAI for systematic benchmarking.