How should real estate agents test Claude Cowork after its Windows fix?

If Claude Cowork stopped running local commands on a Windows computer after the September 8 update, start by checking the official incident page and your Windows version. Anthropic says Microsoft’s September 14 update resolved the issue. After the update is approved and installed through your normal process, test Cowork with a disposable folder and fictional files before letting it touch current listings, client material, or business records.

This was a specific compatibility problem, not evidence that every part of Claude had failed. Anthropic reported that affected Cowork workspaces could no longer reach the computer’s drive, so local commands did not run. For most users, chat plus file reading and editing still worked. Anthropic also said restarting or reinstalling Claude would not fix it while the incident was active.

On September 14, Anthropic marked the incident resolved and linked directly to Microsoft update KB5129195. Microsoft describes that release as an out-of-band cumulative update for Windows 11 versions 24H2 and 25H2. Its fixes include a host-folder access problem affecting some applications that use HCS-managed virtual machines. Microsoft’s page does not name Claude, but Anthropic identifies that update as the fix for Cowork.

First decide whether this incident applies to you

The narrow scope matters. If you use Claude only for ordinary web or desktop chat, this local-command incident may not have affected your work. If you use Cowork on Windows but never saw local-command failures after September 8, do not manufacture a problem to solve. Check the status page, your Windows edition and version, your organization’s update policy, and the behavior you actually observed.

If a brokerage, IT provider, or device-management policy controls updates, follow that process. Microsoft says KB5129195 applies to Windows 11 24H2 and 25H2 and is delivered automatically through Windows Update and Microsoft Update. That does not make a public blog the right authority for your specific computer, security posture, or timing.

Before an operating-system update, save active work and make sure normal backups are current. Record the Windows build and Cowork behavior you have now. A short before-and-after note is more useful than relying on memory later.

Use a small test that can fail safely

A green status page tells you the vendor believes the incident is resolved. It does not prove your own workflow, permissions, or files are ready. Your first test should be small enough to inspect and harmless enough to repeat.

Create a temporary test folder containing fictional material: perhaps three made-up property flyers with invented addresses and a text file listing equally fictional open-house times. Ask Cowork to perform one bounded local task you already understand, such as listing the filenames and creating a separate summary file inside that same test folder. Define the expected result before you run it.

Then inspect the result yourself. Did Cowork reach only the intended folder? Did it create the requested file in the right place? Are the source files unchanged? Did it report an error clearly? A successful chat response is not the test; the local action and its boundaries are.

Do not begin the recovery test with a live transaction folder, downloaded identification, contracts, access instructions, financial records, or a CRM export. Those materials add privacy, security, retention, and professional obligations that a basic compatibility test does not need.

Test the boundary, not just the happy path

After the simple task works, try one controlled boundary check. Ask Cowork to stop before replacing a file, or give it a task that should stay inside the test folder. Confirm that your permissions and approval points behave as expected. If the tool reaches beyond the intended location, produces inconsistent results, or cannot complete the action, stop and capture the exact error instead of repeatedly widening access.

For real estate work, the most important boundary is often the final external action. A tool may help organize approved notes, rename copies of marketing assets, or prepare an internal summary. A person should still verify property facts, deadlines, recipients, and commitments, and should control anything sent to a client, published publicly, changed in a transaction system, or deleted.

Keep a manual route for time-sensitive work

The useful lesson is larger than one Windows patch: an AI workspace can be partly available. Chat may work while local commands fail; a status page may be green while one computer still needs an update. Build workflows so you can see which layer failed and return to a familiar manual method.

For a deadline-sensitive task, know where the original files live, how to complete the job without Cowork, and who should be told if automation is unavailable. Recovery should never require an AI tool to remember the process that replaces it. That would be a very modern little circle.

A short action checklist

• Confirm that you use Claude Cowork on Windows and actually saw the local-command problem.
• Read Anthropic’s incident page and verify your Windows version against Microsoft’s update page.
• Follow brokerage, IT, and device-management rules for operating-system updates.
• Save active work and confirm normal backups before updating.
• Retest with fictional files in a disposable folder, not live client or transaction data.
• Check the resulting files and folder boundaries yourself.
• Keep sends, publishing, system changes, deletions, and consequential decisions under human approval.
• Use the manual process if the test is unclear or fails.

Return to useful work one step at a time

A repaired feature does not need a dramatic relaunch. Prove one local task, inspect it, and expand only when the evidence supports the next step. If you want guided practice designing tests, approval points, and manual fallbacks, join the free AI Agents for Agents Skool community. The deeper templates and implementation lessons live there; the decision rule here is complete: verify the fix on your setup with low-risk material before resuming real work.

Test small. Verify locally.

Build AI workflows that can recover.

Join the free community for practical lessons on bounded tests, human approval, and dependable manual fallbacks.

Join AI Agents for Agents free on Skool →