Copilot Studio’s Computer-Using Agents Go Live: Automate the Apps That Never Had APIs

Every organisation has a handful of systems that quietly run the business and refuse to integrate with anything else. The ageing line-of-business app with no public interface, the supplier portal that only exists in a browser, the desktop tool the finance team still depends on. Until recently, automating work across those systems meant either a brittle script that broke the moment a button moved, or a person copying data between windows by hand. Microsoft’s latest Copilot Studio update takes a direct run at that problem.

What Microsoft announced

Microsoft has made computer-using agents generally available in Copilot Studio. These agents interact directly with websites and desktop applications through the user interface (clicking, typing and navigating the way a person would), so they can automate processes in systems that simply have no API to call. According to Microsoft, the general-availability release adds enterprise capabilities including secure credential management, the ability to select different models for different scenarios, and automations designed to be resilient to interface changes rather than shattering the first time a layout shifts.

Two more capabilities reached general availability alongside it:

• Agent-to-agent (A2A) communication: agents can now exchange information, delegate tasks and be set up to work together across systems and workflows, rather than each operating in isolation.
• Real-time voice agents: available in North America through Dynamics 365 Contact Center, these build more natural voice experiences that can identify callers, answer questions, take action during a conversation, and hand off to a live agent while preserving context.

Microsoft also previewed a redesigned workflows experience, available in early release environments. It offers a more visual designer for orchestrating agentic automation in one place, and lets teams drop existing agents into workflows as nodes, combining deterministic, step-by-step orchestration with the adaptive judgement of an agent.

Why this matters for businesses

The significance is less about a single feature and more about which systems are now reachable. For years, the unwritten rule of automation was that you could only automate what exposed an API. Everything else (and in most established companies, that is a great deal) fell to manual effort or fragile screen-scraping. A computer-using agent moves the line. If a person can do the task through a screen, an agent can be taught to do it too, without waiting for a vendor to ship an integration that may never come.

The move from preview to general availability matters as well. GA is the point at which a capability is meant to be supported and dependable enough to put real work behind, rather than a demo to admire. Paired with A2A communication, the picture shifts from a single clever helper to a set of agents that can coordinate: one reading data out of a legacy app, another acting on it elsewhere. And real-time voice extends the same idea to the phone, where so much customer contact still happens.

A note of realism

Driving an application through its interface is inherently more delicate than calling a clean API, which is precisely why resilience to interface changes is called out as a feature rather than assumed. An agent acting on your behalf inside a real system also holds real permissions, so secure credential management and clear limits are not optional extras. The sensible path is to start with a well-bounded task, watch how the agent behaves, and keep a person in the loop where the stakes are high. Then widen the remit as confidence grows.

What to do about it

The most useful first move is not to buy a tool but to make a list. Which processes in your business depend on systems that never had an API? Where are people moving data between windows, re-keying figures, or working a portal by hand because there was no other way? Those are exactly the tasks this technology is built for, and naming them turns a vague sense of inefficiency into a concrete backlog.

From there, pick one well-scoped task with a measurable outcome, build it properly with credentials handled securely and a human checkpoint where it counts, and measure whether it genuinely saves time and reduces error before you scale. The goal is not to automate everything at once; it is to retire the brittle scripts and manual workarounds one dependable agent at a time.

Source: Microsoft: New and improved computer-using agents, a new workflows experience, and real-time voice experiences