Amazon WorkSpaces for AI agents review: a governed desk for agents that use old software. Pick whose login it borrows
Every company has one program that has no API and eats hours of somebody's week. AWS now offers agents a cloud desktop to work in it, with sign-in and a video record attached. Worth a pilot, if you're careful about whose account the agent uses.

Amazon WorkSpaces for AI agents is worth piloting if you already run on AWS and have a desktop application with no API that staff operate by hand. The November 16, 2026 update added OIDC sign-in, on-behalf-of sign-in so an agent works as a specific user, and session replay stored in your own S3 bucket. If the application has an API, use that instead, and decide whose permissions the agent borrows before it starts.
The short verdict
Worth a pilot if three things are true: you already run on AWS, you have a desktop application with no API that people operate by hand for hours a week, and you can name exactly whose account the agent should use. If the application has an API, use the API. It will be faster, cheaper and easier to test than any agent clicking through screens.
Amazon WorkSpaces for AI agents gives an AI agent its own cloud desktop to work in, the same kind of managed desktop AWS sells for people. The agent sees the screen, clicks, types and scrolls, through a managed MCP server that most agent frameworks can connect to. AWS's own examples are the jobs every operations team recognizes: entering transactions in an ERP and working through a claims queue.
What changed this month
Two updates landed in November. On November 3, AWS added Linux and Windows 11 desktops alongside Windows Server, so an agent can run on whatever operating system the application needs. The Linux desktops come with full root access, which AWS pitches at developer tasks like builds and tests. A November 13 AWS blog post added OAuth sign-in, so agents such as Kiro and Claude Code can connect through a browser login while IAM permissions and CloudTrail logging stay in place.
The November 16 update is the bigger one for anyone thinking about risk. Agents can now sign in through any identity provider that supports OIDC, as well as SAML. With on-behalf-of sign-in, the agent works as a specific signed-in user, "with that user's existing directory identity, permissions, and Group Policies on the desktop." And session replay turns the session's screenshots into a video with a plain-language transcript of what the agent did, stored in your own S3 bucket and encrypted with your own keys.
AWS also made setup simpler. You can describe what the agent needs to an AI coding assistant, and AWS's MCP skills guide it through building the desktop from defaults you can adjust. Teams on Bedrock AgentCore can connect an agent using AWS credentials without writing agent code.
An agent signed in as a person can do everything that person can do, on a bad day too.
What it's good at
The appeal is plain. Plenty of software that runs real businesses was never built to be automated: an old accounting package, a carrier's desktop client, a line-of-business tool that only works on Windows. Until now, giving an agent access meant someone building a fragile setup on a spare machine and hoping security never asked about it.
WorkSpaces puts that work inside controls a security team already knows: network isolation, IAM access, corporate sign-in, and a record of what happened. Session replay is the part we'd value most. When an agent does something odd at 2 a.m., a video and a transcript of what it saw and did is the difference between a five-minute answer and a week of guessing.
Decision guide
What AWS now handles, and what stays with you
| Consideration | AWS now provides | Still yours to decide |
|---|---|---|
| Desk | A cloud desktop on Windows Server, Windows 11 or Linux | Which applications and drives that desktop can reach |
| Identity | OIDC or SAML sign-in, and on-behalf-of a user | Which user it borrows, and what that user is allowed to do |
| Record | Session replay video and transcript in your S3 bucket | Who watches it, how often, and what counts as a problem |
| Setup | AI coding assistants can build the WorkSpace from defaults | Reviewing those defaults before anything real runs |
Where it falls short
An agent signed in as a person can do everything that person can do, on a bad day too. On-behalf-of sign-in is honest about identity, which is good. It also means the agent inherits that user's permissions, all of them. If the account can approve payments, so can the agent. The fix is old-fashioned: create a dedicated account with only the access the job needs, and give it a human owner.
A replay is a record, not a guardrail. It tells you what happened after the fact. It doesn't stop the agent from posting the wrong invoice. You still need limits inside the application, approval steps for anything that moves money or data out, and a habit of actually watching a sample of sessions.
Screens change. Agents that work through the interface are sensitive to a moved button or a new pop-up in a way an API integration isn't. Plan for a test run after every update to the application.
Licensing is the question teams forget. Many desktop applications are licensed per named user or per seat, and the license terms may say something about automated access. A dedicated agent account may need its own seat, and the vendor may have a view on software operating it. Read the agreement, or ask the vendor, before the pilot rather than after it works.
Then there's the question of what happens when the agent gets stuck. A pop-up it hasn't seen, a login that expired, a record locked by someone else: decide in advance whether the agent stops and tells someone, retries, or skips the item, and make sure a person actually gets the message.
Pricing deserves a careful look. AWS's announcements don't publish a separate price for agents. The setup guide warns that the stacks and fleets you create may result in charges until you delete or stop them, so treat it like any other AWS desktop resource and price the hours the agent will actually run. Availability also depends on Region: AWS says the features are live where the WorkSpaces Applications MCP endpoint is offered.
Against the alternatives
Compared with an API integration, it's slower and more brittle, but it works where there's no API, which is the whole point. Compared with a homemade setup, an agent on a spare PC or a virtual machine nobody manages, it's clearly better on identity, logging and isolation. Compared with consumer agents that use their own cloud computers, it's built for your systems and your accounts rather than the public web.
Sequence of work
How the product got here
- Earlier in 2026
WorkSpaces for AI agents launches on Windows Server desktops.
- Nov 3 and Nov 13
Linux and Windows 11 desktops added; OAuth browser sign-in for MCP agents.
- Nov 16
OIDC and on-behalf-of sign-in, session replay, and setup through coding assistants.
- Before your pilot
Check your Region has the WorkSpaces Applications MCP endpoint.
Our verdict
Pick one application with no API, one repetitive job and one dedicated account with the narrowest permissions that job needs. Run the pilot with a person reviewing every session replay for the first two weeks, and write down what "done correctly" looks like before the agent starts. If it holds up, widen it slowly.
Measure it against the person who did the job before: how long each item took, how many needed correcting, and how much time went into reviewing the agent's work. If the review takes as long as the job, the pilot has told you something useful.
Scoping that pilot, building the agent, and deciding what it may do without asking is the work of our AI applications and agents practice. Talk to us.
Related work: Spin:Market
We build alongside the people doing the work.
Bring us the problem you’re thinking about.