Copilot Studio Computer-Using Agents Are Now Generally Available — Plus New Workflows and Voice Updates
Microsoft's Copilot Studio CUA feature hits GA, letting agents operate any UI like a human, with new workflow tools and real-time voice now live too.
Microsoft made computer-using agents (CUA) in Copilot Studio generally available on 13 May 2026, rolling out across all commercial Power Platform geographies. The same release window brought a redesigned workflows experience into early release and pushed real-time voice agents to GA in North America. Taken together, this is a meaningful step away from AI that answers questions and toward AI that actually does things.
What Computer-Using Agents Actually Do
The concept is straightforward, even if the technology behind it is not. A computer-using agent gets the same tools a person has: a browser, a screen, a keyboard, and the ability to read what is on the page and decide what to do next. Rather than relying on brittle selector-based automation that breaks whenever a UI shifts, CUA uses vision and reasoning to navigate live interfaces, adapt to layout changes, and complete multi-step tasks without a human clicking through each screen.
To set one up, you create or open an agent in Copilot Studio, go to Tools, add a new computer use tool, and describe the task in plain language. That is genuinely the starting point.
The GA build ships with support for both OpenAI CUA and Anthropic Claude Sonnet 4.5 models, Azure Key Vault for credential storage, Microsoft Purview audit logging, configurable human-in-the-loop review, and Windows 365 Cloud PC pool support. There are real limits worth noting: sovereign clouds (GCC, GCC High, DoD) are not in this rollout, and Electron, Java, Unity, Citrix, and most virtualised environments are not supported yet.
What This Means for Enterprise IT Teams
For organisations sitting on a stack of legacy systems that were never built with APIs in mind, this changes the calculus on automation significantly. Workflows that previously required either a months-long integration project or people manually clicking through proprietary portals can now be handed to an agent.
A real example from the release: Graebel, working with GET AI and Microsoft, built a Service Order Agent that monitors mailboxes, extracts data from unstructured service-order emails using Azure Content Understanding, validates each request against business rules and compliance requirements, and then operates their Global Connect system directly through its UI, navigating screens and entering data exactly as a trained operator would. No API, no platform redevelopment. Exceptions and low-confidence cases get routed to humans before any action is taken.
That last part matters. The human-in-the-loop capability is now in preview as a first-class feature, using a “request information” action that pauses an agent flow, sends the relevant details to a designated reviewer via Outlook, and resumes execution using their response. It is a practical answer to the reasonable concern that autonomous agents will just barrel ahead when they should stop.
The Redesigned Workflows Experience
The new workflows canvas is in early release environments now. It brings a unified visual designer where you can build end-to-end automation, dropping in agent nodes alongside API calls, approvals, and business logic on a single canvas.
Previously, you would stitch together disconnected tools across multiple surfaces and hope the seams held. The redesign puts AI-powered steps, including classification, content generation, and decision support, directly into the same orchestration layer as deterministic workflow logic. Structured where the process demands it, adaptive where it helps.
Computer-using agents can now be embedded directly into these multi-step workflows, which is currently moving into preview. So an agent that navigates a legacy UI can be one step in a broader business process, not a separate system running in isolation.
Worth flagging on the performance side: Microsoft’s updated orchestration layer improves evaluation performance by roughly 20 percent while cutting net token consumption by 50 percent, based on their 2026 usage data. Fewer tokens means lower cost per workflow run, which adds up quickly at scale.
Real-Time Voice Agents
Real-time voice agents are now generally available in North America through Dynamics 365 Contact Center, with additional regions and language support planned as part of the broader rollout.
For contact centre teams, the pitch is replacing rigid phone trees with agents that can identify callers, answer questions, and take action during the call itself. Microsoft has published external voice agent templates covering common customer service workflows including billing and payments, which gives teams a starting point rather than a blank canvas.
Over 80 percent of Fortune 500 companies have active agents built in Copilot Studio, so there is a large base of organisations that can now layer voice capability directly into existing agent-based workflows rather than standing up something separate.
Agent-to-Agent Communication and Work IQ
One quieter but significant addition: agent-to-agent (A2A) communication is now generally available in Copilot Studio. Agents can exchange information, delegate tasks, and coordinate across systems. Combined with new REST API access, CLI support, and remote Model Context Protocol server support, the architecture for building networks of specialised agents is becoming more concrete.
If you are building or maintaining Power Platform integrations, there is also a namespace change to be aware of: update API calls to use the new copilotstudio namespace. The old namespace continues to work for now, but switching early keeps you ahead of any future deprecation.
The Practical Takeaway
The preview-to-GA journey for computer-using agents took roughly eight months from the September 2025 preview launch. The GA build is more capable, more governed, and more integrated than the preview. If you explored CUA earlier and held off, the addition of Purview logging, Key Vault credentials, and human-in-the-loop routing addresses most of the governance concerns that would have blocked enterprise adoption.
If you have not looked at it yet, the Graebel example is a useful frame: the value is not replacing humans, it is removing the manual, repetitive UI work that slows humans down, while keeping them in the loop for anything that needs judgement.
The full release details are on the Microsoft Copilot Blog, and the Microsoft Learn what’s new page is kept current as features roll out.