Microsoft Copilot Studio agents go GA

Microsoft said on May 13 that computer-using agents in Copilot Studio are now generally available, giving enterprise customers a way to let AI agents click through websites and desktop applications, enter text and complete UI-based tasks. The release extends the product beyond drafting and chat-style assistance into workflow execution inside systems that may not expose APIs. (techcommunity.microsoft.com) Microsoft said the feature is being expanded to all commercial geographies in Microsoft Power Platform, with deployment under a tenant’s data residency and compliance boundaries. The company framed the launch as an answer to automation work that has remained in vendor portals, internal web apps and line-of-business systems. ### What exactly became generally available on May 13? Microsoft’s May 13 post said “computer use” in Copilot Studio is now generally available, allowing makers to add a tool that lets an agent operate software through a graphical interface. (techcommunity.microsoft.com) The agent can select buttons, choose menus and type into fields using a virtual mouse and keyboard, according to Microsoft’s documentation. The product documentation says the tool works with websites and desktop apps on Windows computers and can be described in natural language rather than code. Microsoft said that matters when a company needs to automate tasks in systems that do not provide direct API access. ### How is Microsoft describing the technical change? (techcommunity.microsoft.com) Microsoft Learn says computer use is powered by a Computer-Using Agents model that combines vision and reasoning to interact with graphical user interfaces. The documentation says the tool is designed to adapt when buttons move or screens change, which Microsoft presents as a way to reduce breakage from interface updates. Mustapha Lazrek, writing in Microsoft’s Copilot Studio blog, said the release is aimed at “long-tail, UI-driven business processes” that previously required robotic process automation scripts or manual work. (techcommunity.microsoft.com) Microsoft said every Copilot Studio maker can now build agents that “take action directly inside any application a person can use.” (learn.microsoft.com) ### Where do Agent Skills fit into this release? Microsoft’s skills documentation says Copilot Studio agents can be extended with skills built using the Microsoft 365 Agents SDK or other pro-code tools and then registered inside a Copilot Studio agent. Once registered, those skills can be triggered in conversation, according to Microsoft Learn. The same documentation draws a line between flows and skills. (learn.microsoft.com) Microsoft says flows are suited to simpler single-turn operations such as placing an order or checking status, while skills are intended for more complex, multi-turn actions such as scheduling a meeting or booking a flight. That gives companies a way to package repeatable behavior and attach it to broader agent workflows. ### What controls did Microsoft add around agent actions? (techcommunity.microsoft.com) Microsoft’s May 13 post said the generally available release includes built-in credentials and Azure Key Vault support for signing in to websites or desktop applications. The company also said the feature includes allow lists for websites or desktop applications, Power Platform data loss prevention policies, environment isolation and audit trails. (learn.microsoft.com) The same post said human-in-the-loop checkpoints can be inserted for low-confidence steps, exceptions and decisions that require operator approval. Microsoft also said run history shows what the agent saw, clicked and why, and that logs are propagated to Microsoft Purview and Dataverse for audit and administrative review. (learn.microsoft.com) ### What does Microsoft say about model choice and availability? Microsoft’s documentation says supported models for computer use currently include OpenAI’s Computer-Using Agent and Anthropic’s Claude Sonnet 4.5. The documentation adds that Anthropic’s model requires an administrator to enable external models. Microsoft Learn also says the feature is available for environments where the region is set to the United States. (techcommunity.microsoft.com) Microsoft’s May 13 announcement, by contrast, said the company is expanding availability to all commercial Power Platform geographies. The company’s public materials therefore point to a rollout that is broader at the platform level while some setup requirements remain spelled out in current documentation. ### How are teams supposed to measure whether these agents are working? Microsoft said in a May 11 update that Copilot Studio now surfaces agent status in the authoring experience and offers an Analytics Viewer role that is generally available for read-only access to analytics. Mohamed Arhab, a solution architect at the City of Montreal, said the role lets stakeholders monitor performance while keeping configuration and publishing controls separate. (learn.microsoft.com) A separate Microsoft post said Agent Evaluation in Copilot Studio is now generally available and gives teams a no-code way to test and monitor agent quality, safety and related measures. That gives customers a second layer of instrumentation alongside run history and audit logs as they move from pilot use into production workflows. (techcommunity.microsoft.com) Microsoft’s next published checkpoints are already on its public sites: the Copilot Studio monthly updates page dated May 11, 2026, the May 13 general-availability post by Mustapha Lazrek, and the Microsoft Learn pages covering computer use, skills and configuration. (techcommunity.microsoft.com 1) (techcommunity.microsoft.com 2) (microsoft.com)