The Pulse
Windows 11 records coding-agent use in diagnostics, researcher says
Xusheng Li found that a Windows 11 component records which coding agents and developer tools have been observed. In his tests, the resulting report entered an upload queue when optional diagnostic data was enabled.

A Windows 11 component records which coding agents and developer tools have been used and can queue that information for upload, according to an investigation by Xusheng Li. With optional diagnostic data enabled, Li found the agent-usage report in the Windows telemetry service's upload queue.
Li found no prompt text or source code in the report he examined. The component was active in his virtual machine, but he said his evidence did not establish that it is active on every Windows installation.
What the component checks
Li's investigation focused on MissionControlAggregator.dll, a Windows component containing checks against executable names. He found four command-line agent slots that check for copilot.exe, claude.exe, codex.exe and agy.exe, including extensionless equivalents on WSL route. Five desktop-agent slots check github.exe, claude.exe, codex.exe, cursor.exe and antigravity.exe.
The component also checks names associated with developer tools, including node.exe, python.exe, dotnet.exe, gcc.exe and code.exe, and groups them into broader categories. Some developer-tool checks also verify the software publisher; the compiler and debugger group containing GCC and Clang has no additional publisher check in the path Li examined.
Li traced the inputs to existing Windows events rather than a process scan performed by the component itself. Application-state and interactivity summaries supply identities for Windows programs, while separate events supply image names for programs running through WSL.
From a match to a report
A successful match sets a slot in the component's state, which is saved locally. The component later turns the slots into bitsets in a diagnostic event called Segments.SubCohorts, Li reported. A set bit indicates that the corresponding tool was observed during the reporting window; executable names are not written into each report.
The source events contain details such as runtime, time in focus, interaction counters and resource usage. The component appears to discard those details and retain a bitmap of observed agents, according to Li. The report's timestamps describe the collection window rather than the duration of each agent's use.
Li examined a Parallels Windows 11 ARM64 virtual machine running build 26200.9974. Its coding-agent feature was enabled as a service experiment before his test, he said, and the Connected User Experiences and Telemetry service was running automatically. That service starts AggregatorHost.exe, which loads the component.
For a practical test, Li copied cmd.exe to a file named codex.exe and ran it in an interactive Windows session. The name alone was enough for the component to recognize the file as Codex, he found; the corresponding bit then appeared in an emitted report. A separate test using a copy named gcc.exe set a developer-tool slot shared by the compiler and debugger category.
When data enters the queue
The report normally follows a roughly daily boundary. To capture it immediately, Li moved the saved reporting-window start 25 hours into the past and restarted the diagnostics service while listening for the event.
Finding a local diagnostic event alone would not establish that it was sent out. Li then inspected the service's stored upload batches and found the agent event in the queue when optional diagnostic data was allowed. He also found the developer-tool event in the queue under that condition.
In a clean installation of Windows 11 Home ARM64 with the region set to the United States, optional diagnostic data was enabled before he changed any privacy settings, Li reported. That observation was limited to the US setup he tested. Li said users can decline optional diagnostics during setup or turn off the setting later under Windows' diagnostics and feedback settings.
Source
Related stories
- Hamburg regulator issues privacy checklist for AI glasses influencers
- Plugin4Shell Breaks SHA Pinning in Four AI Coding Agents
- NameTrap Shows How Variable Names Can Steer Coding Agents Into Bugs
- Fake Apex Accelerator Routed AI Coding Traffic to Its Servers
- MAGS Verifies Agent-Written Code, Then Exposes Its Limits