Microsoft announced Copilot Cowork on March 9, 2026, an autonomous agent for Microsoft 365 Copilot built with technology from Anthropic's Claude Cowork. Cowork runs in the cloud inside Microsoft 365, breaks a request into steps and carries them out across Outlook, Teams, Word, Excel, PowerPoint and other apps, pausing for approval before important actions such as sending an email. For developers and architects, it is a high-profile example of an enterprise agent design, and further evidence that Microsoft's AI strategy now spans multiple model providers.

Key Facts#

  • Announcement: March 9, 2026, as reported by VentureBeat and GeekWire.
  • Anthropic inside: Microsoft's FAQ says Cowork uses Anthropic models, with Anthropic acting as a Microsoft subprocessor.
  • Execution: tasks run in the cloud in an isolated environment, and Cowork is available in the browser and in Microsoft 365 Copilot desktop apps.
  • Reach: Outlook, Word, Excel, PowerPoint, Teams, SharePoint, OneDrive and PDF files.
  • Human oversight: Cowork asks for approval before sensitive actions, and users can pause, resume or cancel tasks.
  • Skills: 13 built-in skills, plus custom skills written as Markdown files stored in OneDrive.
  • Licensing: GeekWire reported that the launch came with a new Microsoft 365 E7 licensing tier.

What Happened#

Microsoft introduced Cowork as a way for Microsoft 365 Copilot to do work rather than only answer questions. According to Microsoft's documentation, users describe a task in natural language, Cowork breaks it into steps that it shows in the conversation, and then it drafts emails, schedules meetings, creates documents, posts in Teams, searches organizational resources and manages files. Users can interrupt and redirect it at any point. Before medium- and high-risk actions, such as sending a message or scheduling a meeting, Cowork pauses and asks for a go-ahead.

The agent's behavior is extended through skills. Cowork ships with 13 built-in skills covering areas such as Word, Excel, PowerPoint, email, scheduling and enterprise search, and users can add their own by placing Markdown files in a OneDrive folder. An April 2026 update raised the limit on custom skills from 20 to 50 per user, added support for Markdown skill files in OneDrive and SharePoint, and introduced searchable task history and in-app notifications.

The Anthropic connection is explicit. Coverage of the launch described Cowork as bringing the technology behind Anthropic's Claude Cowork into Microsoft 365, and Microsoft's own FAQ states that Cowork uses Anthropic models as a subprocessor. VentureBeat characterized it as a cloud-powered agent that works across Microsoft 365 apps, coverage noted that it draws on a user's enterprise work data, and GeekWire tied the rollout to a new E7 licensing tier.

Background#

According to launch coverage, Anthropic released Claude Cowork as a research preview for macOS in January 2026 and followed with a Windows version in February. Microsoft had already been integrating Anthropic models into its products. At Ignite in November 2025, it added Claude models to Microsoft Foundry, and its documentation on Anthropic as a subprocessor describes Claude access in Microsoft 365 Copilot, Researcher, Copilot Studio, Agent Mode in Excel and the Word, Excel and PowerPoint agents, all covered by Microsoft's Product Terms and Data Protection Addendum.

That documentation also records the rollout mechanics. A new admin toggle for Anthropic models appeared in the Microsoft 365 admin center on December 8, 2025, enabled by default for most commercial customers, and the older Anthropic commercial terms option was deprecated on January 7, 2026. There are important exceptions: Anthropic models are currently excluded from the EU Data Boundary, the toggle defaults to off for EU, EFTA and UK customers, and the models are not available in government or sovereign clouds.

Why It Matters for Developers#

Most developers will not build Cowork, but many will build agents that look like it, and its design choices are worth studying.

  • Approval gates by risk level. Cowork distinguishes routine steps from actions with side effects and asks for confirmation only when it matters. That is the right default for any agent that can send messages or change data. Our AI agent architecture patterns guide covers human-in-the-loop designs in .NET.
  • Skills as plain Markdown. Letting users extend an agent with Markdown instructions lowers the barrier to customization, but it also means instructions can come from many authors. Treat skill content like any other input that can be manipulated, and apply the defenses in our responsible AI and LLM security guide.
  • Permissions and data exposure. An agent that searches across mail, files and chats will surface whatever a user can access. Organizations should review oversharing in SharePoint and Teams before broad rollout, and developers building Microsoft 365 integrations should expect agents to read and act on the data their apps create.
  • Data residency is model-specific. Because Cowork relies on Anthropic models, the EU Data Boundary exclusion and regional defaults in Microsoft's subprocessor documentation are directly relevant to compliance reviews.

The launch also reinforces a strategic point. Microsoft now ships products built on OpenAI models, its own MAI models and Anthropic models. For .NET teams building on Microsoft Foundry, the same multi-provider reality argues for keeping model access behind abstractions such as IChatClient from Microsoft.Extensions.AI, so the choice of model can change without rewriting the application.

What's Next#

Cowork entered the Frontier program for early access shortly after the announcement, and later reports said it reached general availability in June 2026. The larger questions concern pricing and packaging under the new E7 tier, whether Anthropic models will come inside the EU Data Boundary, and how Microsoft will balance Anthropic, OpenAI and in-house models across Copilot. The April 2026 amendment to the Microsoft–OpenAI partnership, covered in our story on the new terms, made that multi-model direction even more explicit.

Sources#