Microsoft Ships Computer-Using Agents in Copilot Studio to GA — AI That Drives Any App, No API Needed

Microsoft has made computer-using agents in Copilot Studio generally available, rolling the capability out across every commercial Power Platform region and becoming the first major hyperscaler to put production-grade "computer use" into customers' hands. Announced on May 13 via the Copilot Studio blog, the launch moves a capability that has mostly lived in research demos into something enterprises can deploy under their own tenant and compliance boundaries.

The pitch is simple and, for a lot of back-office work, overdue: instead of wiring up integrations, an agent can operate desktop and web applications the way a person does — reading the screen, moving the cursor, clicking buttons, filling forms and navigating between windows. That lets organizations automate processes sitting on top of legacy or third-party software that never exposed an API, the exact cases where brittle robotic-process-automation scripts tend to break whenever a vendor nudges its UI.

Crucially, Microsoft is shipping this as an enterprise product, not a tech demo. The GA build runs on OpenAI's computer-using-agent model and Anthropic's Claude Sonnet 4.5, and wraps them in the governance plumbing IT departments demand: Azure Key Vault for credential storage, Microsoft Purview audit logging of what the agent did, and configurable human-in-the-loop review so a person can approve sensitive steps before the agent proceeds.

The timing matters competitively. As of late May, the same capability from rivals is still earlier on the maturity curve — Anthropic's computer use sits in a paid beta and Google's Gemini Computer Use is in public preview — which lets Microsoft claim the first full general availability among the big platforms. It also reframes the race: the question is shifting from which lab has the smartest model to who can wrap computer-use models in the security, auditing and reliability that turn a flashy capability into something a Fortune 500 will actually run in production.

That has implications well beyond Copilot. Production-grade computer use puts direct pressure on standalone RPA vendors whose moat was exactly this kind of UI automation, and it nudges the broader "agent economy" from chat-style assistants toward software that takes real actions inside other software. The flip side is risk: an agent that can click anything a user can click is only as safe as its guardrails — which is why the audit trail and human-in-the-loop controls, not the demo of an agent booking a flight, are the part Microsoft is keenest to talk about.