AI Copilots Becoming Work Agents

AI copilots are being rebuilt as workplace agents that can search company data, act inside apps, and run tasks on a schedule. (help.openai.com) OpenAI said this week that ChatGPT workspace agents are rolling out to Business and Enterprise customers over the next few weeks. The company says the agents can be built from templates or from scratch, run in ChatGPT or Slack, and connect to Google Drive, Google Calendar, Slack, SharePoint, files, skills, and custom Model Context Protocol servers. (help.openai.com) OpenAI’s help center says workspace agents can be previewed before publishing, shared inside a workspace, and run on schedules for repeatable workflows. The same documentation ties the feature to managed workspaces with member roles and permissions rather than to individual consumer chats. (help.openai.com 1) (help.openai.com 2) Anthropic moved in a parallel direction on April 17, when it added Claude Design in research preview alongside Claude Opus 4.7. Anthropic says Claude Design can generate visual designs, interactive prototypes, presentations, and one-pagers through chat, and it is available to Pro, Max, Team, and Enterprise plans, with Enterprise access turned off by default. (support.claude.com 1) (support.claude.com 2) (support.claude.com 3) That default-off setting is part of the same shift. Anthropic’s design-system guide says admins can ground Claude Design in codebases, slide decks, and other brand assets so outputs follow a company’s own components, colors, typography, and patterns. (support.claude.com) At the same time, Anthropic has been testing tighter limits around higher-risk coding tools. The Register reported on April 22 that Anthropic removed Claude Code from some public Pro-plan pages, and Anthropic told the publication the change was a test affecting about 2% of new users. (theregister.com) The reason those controls are getting tighter is visible in Mozilla’s latest Firefox release. Mozilla said on April 21 that Firefox 150 shipped fixes for 271 vulnerabilities found during an evaluation of Anthropic’s early Claude Mythos Preview model. (blog.mozilla.org) Mozilla said the same collaboration had earlier used Claude Opus 4.6 to identify 22 vulnerabilities in Firefox 148, a much smaller count than the 271 fixes now landing in Firefox 150. Ars Technica reported that Anthropic is limiting Mythos Preview to a small group of critical industry partners while it studies the model’s offensive security capability. (blog.mozilla.org) (arstechnica.com) The product pattern is getting clearer across vendors: broader access for agents inside managed workspaces, narrower access for tools that can write code or uncover exploits at expert level. OpenAI’s rollout emphasizes sharing, previews, schedules, and workspace controls, while Anthropic is pairing new creative and coding features with previews, admin defaults, and restricted release channels. (help.openai.com 1) (help.openai.com 2) (support.claude.com) (theregister.com) (blog.mozilla.org) The result is that “copilot” now means less a chat box that suggests text and more a governed worker inside the company stack. The next fight is likely to center on who can grant those agents access, what they are allowed to touch, and how every action gets logged after they do it. (help.openai.com) (help.openai.com) (support.claude.com)