Microsoft gives AI hands with Computer-Using Agent in Azure

Automating real-world work with AI has always hit an annoying wall. If an application lacks a clean API, developers are stuck writing brittle RPA scripts that snap the moment a developer shifts a button two pixels to the left.

Microsoft announced two major additions to Azure AI Foundry to fix this: a specialized Computer-Using Agent (CUA) model that can look at screens and control software, and the Responses API, designed to simplify how developers call multiple agent tools in one go.

Browsing and clicking through legacy software

CUA is a visual agent model built inside Azure OpenAI Service. Instead of relying on APIs, it looks at the graphical user interface just like a human does, recognizing buttons, filling out forms, and navigating across both desktop and web applications using natural language prompts.

Because giving an autonomous agent access to a local mouse and keyboard is inherently risky, Microsoft is evaluating integration with Windows 365 Cloud PCs and Azure Virtual Desktop. That would keep the agent sandboxed inside managed virtual machines rather than messing around on an employee's physical machine.

Microsoft also acknowledged the current technical limitations. The company explicitly noted that CUA’s reliability in non-browser environments is still maturing. As a result, the model is configured to pause and request human confirmation for irreversible steps, while blocking high-risk operations like financial transactions outright.

One API call to rule all agent tools

Alongside the model, Microsoft introduced the Responses API. Until now, Azure developers often had to choose between the lightweight Chat Completions API and the more complex Assistants API.

Responses API merges the best of both. Developers can now execute file searches, custom function calling, and CUA computer use within a single API request.

It also introduces unique response IDs, allowing systems to chain multiple responses into coherent multi-turn conversations without forcing developers to manually build tedious state-tracking pipelines.

Why it matters

In our view, Microsoft’s real masterstroke here isn't just matching the computer-use capabilities first popularized by Anthropic—it is leveraging the Azure and Windows enterprise moat.

Spinning up computer-use agents inside managed Windows 365 environments solves the biggest blocker holding back enterprise adoption: security and compliance. Companies desperately want to automate legacy back-office tools that will never get native APIs, but they are terrified of letting autonomous agents run wild on company laptops.

For developers, Responses API significantly lowers the engineering overhead of orchestrating complex agent behaviors. That said, it is not time to hand over critical financial pipelines to CUA just yet; Microsoft’s own warnings make it clear that human supervision remains mandatory for delicate tasks.

Looking at the broader industry, the battle lines among hyperscalers have officially shifted. The race is no longer just about raw reasoning benchmarks, but about who can safely turn AI models into active digital workers inside an enterprise-grade sandbox.