Loading...
Loading...
OpenAI and other AI model makers are pushing deeper integration of large models into developer workflows, turning ad-hoc tasks into reproducible, automatable processes. OpenAI’s Codex Mac update adds Appshots — a quick-press capture that sends screenshots plus extracted off-screen text and hidden UI elements to the model — plus a promoted /goal command for persistent, multi-hour/day tasks across apps, faster browsing, plugin sharing and expanded enterprise analytics. Around the same time, users report models like Qwen3.6 enabling pipeline-driven automation where assistants document reusable “skills” and other models execute them, shifting how people interact with their computers and offloading repeatable devops, testing and content-conversion work to AI.
Deeper model integration automates repetitive developer tasks and bridges GUI interactions with programmatic workflows, increasing productivity and changing toolchains. Tech professionals must adapt architectures, security postures, and CI/CD practices to accommodate agent-driven operations.
Dossier last updated: 2026-05-30 00:23:17
OpenAI expanded Codex remote-control capabilities to Windows 10 and Windows 11, letting users start and monitor Codex tasks on PCs from iPhone and Android ChatGPT apps. The update adds a “computer use” feature that allows Codex to interact with desktop applications by viewing the screen, clicking UI elements, and entering text — automating repetitive foreground tasks without manual intervention. The change brings parity with earlier Mac remote-control support and could streamline mobile-to-PC workflows for automation, testing, and remote assistance, while raising questions about security, consent, and safe automation boundaries.
OpenAI 今天宣布，Codex 的 Computer Use 功能正式登陆 Windows。 这个功能让 AI 能像人一样操作桌面应用，看屏幕、点鼠标、打字，4 月中旬上线时只支持 macOS。 同时更新的还有手机远程控制。5 月中旬 OpenAI 把 Codex 接入了 ChatGPT 手机 App，可以在手机上启动、监控和审批电脑上跑着的 Codex 任务，但当时只能连 Mac 主机。现在 Windows 也能当主机了，出门在外用手机盯着家里的 Windows 电脑干活，流程算是跑通了。
OpenAI has integrated desktop agent features into macOS by leveraging accessibility APIs and a 2025 acquisition. The company bought a small team, Software Applications Incorporated (makers of Sky), whose founders had prior success selling Workflow to Apple; their Sky-based binary powers Codex’s Computer Use capability (SkyComputerUseClient). Mac apps expose rich accessibility trees by default—thanks to Apple’s historical decisions—allowing agents to read off-screen content and operate with an independent virtual mouse without interrupting users. That combination of accessible-by-default UI frameworks and agent-focused tooling helps explain why macOS is currently better positioned for AI-driven desktop automation, a shift that matters for app design, user privacy, and platform strategy.
OpenAI updated its Codex desktop app for Mac with a new Appshots feature that lets users press both Command keys (customizable) to capture the current app window and send not only a screenshot but extracted, off-screen text and hidden UI content to Codex. The change aims to streamline debugging, UI implementation, and complex-interface sharing compared with manual screenshots or copy-paste. OpenAI also graduated the /goal command from experimental to production, enabling persistent, multi-hour or multi-day tasks across the Codex app, IDE extensions, and CLI with progress monitoring and pausing. Additional updates include a faster, more accurate built-in browser with batch commenting, team plugin sharing for Business accounts, enterprise plugin trial access, and expanded analytics for Business and Enterprise admins.
The author says Qwen3.6 35Ba3 has reshaped their workflow by automating complex tasks through a pipeline: they prompt an assistant (Codex) to perform tasks and document the process into a reusable "skill," feed that to a proxy (pi), and let Qwen3.6 execute the heavy lifting. Key uses include VPS DevOps, converting PDFs to EPUBs with Docling, automated testing with Playwright, and completing code tickets. This approach offloads repeatable, error-prone work to the model and changes how the author interacts with their computer, turning workflows into transferable artifacts. It matters because it illustrates practical, productivity-driven integration of large models into everyday developer and ops tasks.