Desktop and Application Streaming

WorkSpaces for AI Agents now available on WorkSpaces Applications image builders

Building an AI agent that operates desktop applications includes testing it against your specific applications and desktop configurations. Often, the first goal is a pilot that proves the agent can run your business workload, and you want to reach that point quickly.

Starting today, WorkSpaces for AI Agents is available on WorkSpaces Applications image builders. You can test and refine your agent while you install and configure your desktop, before you create an image, fleet or stack. This gets you to a working build sooner and shortens the time to value. Once your agent works the way you want, you can move to production with confidence.

What you get with WorkSpaces for AI Agents on image builders

On an image builder, your agent connects to the same managed Model Context Protocol (MCP) server it uses for stacks. The connection and authorization model, the computer-use tools, the security guardrails and the monitoring are all the same. An agent that works on a stack needs little or no change to work on an image builder.

Image builders add two features that help with testing.

First, you choose which Windows user the agent signs in as. An image builder has three built-in Windows users, and the agent names one when it connects. For testing, you usually want TEST_USER, an unprivileged standard account. It shows whether your agent can complete its workflow as a typical end user, not only as an administrator.

Second, the agent can switch users without starting a new session. Two new tools, switch_user and get_current_user, let the agent move between the built-in users during a single run. For example, the agent can connect as ADMINISTRATOR to set up the desktop, switch to TEST_USER, then run its workflow, all without reconnecting.

How it works

The goal is simple: validate your agent’s workflow on the image builder, refine it, and run again.

  1. Turn on WorkSpaces for AI Agents. Enable it when you create an image builder, or add it to one you already use.
  2. Connect your agent. Your agent connects to the image builder and signs in. If it needs to set up the desktop first, it signs in as ADMINISTRATOR and then switches to TEST_USER.
  3. Run your agent’s workflow. Have your agent run the workflow you built for it, such as opening an application, signing in to the application, creating a document and saving it. The agent reads the screen, does the work and reports back. You see how it handles your desktop, including unexpected dialogs and controls that have moved, and you can watch it work in real time from a browser.
  4. Validate, refine and run again. Use the screenshots and audit trail to check that your agent completed its workflow as expected. To refine it, change one thing, such as the prompt, a skill, the model, an application version or a desktop setting, and run again on the same image builder. You can see the effect of each change on its own. When your agent performs reliably, capture your image and move to a stack for production.

Get started

WorkSpaces for AI Agents on image builders gives you a governed environment to test and refine your agents against the applications and desktop configurations you support, before you roll out to production.

To learn more about Amazon WorkSpaces and WorkSpaces for AI Agents, visit Amazon WorkSpaces for AI Agents.

Joe Jabbour is a Senior Specialist Solutions Architect at Amazon Web Services (AWS) focused on End User Computing Services.
Robert Fountain is a Senior Agentic WorkSpaces Specialist SA at AWS with nearly 20 years of experience in end user computing. He joined AWS in 2020. When he isn’t working, he patrols with the National Ski Patrol and spends time with his family.