Skip to content

Transmission 013

Windows 11 Is Quietly Counting Which AI Coding Agents You Run

Tech News5 min read

Signal
Channel
Tech News
Received
Length
5 min read

Plenty of us still keep at least one Windows box around for scripting and dev work, so when a post titled “Windows 11 is collecting data on which coding agents and developer tools you use” shot to the top of Hacker News this week, I read it twice. Then I read the follow-up coverage. Here is what the research actually shows, what it does not show, and why I think it matters more than the bitmap at the center of it.

What the researcher found

Reverse engineer Xusheng Li published his findings on November 16. He went digging through Windows binaries for undocumented features and landed on a component called MissionControlAggregator.dll. It runs inside an AggregatorHost.exe process launched by DiagTrack, the “Connected User Experiences and Telemetry” service most of us have been eyeing suspiciously for years.

Behind a feature flag named CLICodingAgentUsage, the DLL listens to diagnostic events Windows already generates about running apps, including WSL processes, and compares executable names against built-in lists. According to Li, those lists include:

  • Command-line agents: codex, claude, copilot and agy.
  • Desktop agents: GitHub, Claude, Codex, Cursor and Antigravity apps.
  • Developer tool categories: things like node, python, dotnet, gcc, clang, javac, Visual Studio and VS Code.

Each match flips a bit in a small bitmap covering a reporting window of roughly a day. Li then found the resulting event sitting in DiagTrack’s local upload queue, flagged as optional diagnostic data. His proof of concept was refreshingly simple: copy cmd.exe, rename it codex.exe, run it, and the Codex bit lights up. Detection is purely by file name.

What it does not show

Before anyone grabs a pitchfork, the limits matter, and Li is careful about them himself:

  • No prompts, no code. He found no prompt text, source code or file contents in the report. The source events carry details like runtime and focus time, but the component appears to throw those away and keep only yes/no flags.
  • Not confirmed on every PC. On his main test VM (a Windows 11 ARM64 build 26200 machine) the feature had been switched on remotely as a service experiment. A separate VM built from the public ISO had it disabled. How widely Microsoft has rolled it out is unknown.
  • Optional data only. The event is tagged as optional, not required, diagnostic data. That said, on a clean Windows 11 Home install in the US, Li saw “Include Optional” already selected during setup, so plenty of people may have it on without realizing.
  • No Microsoft explanation yet. So far, it appears Microsoft has been quiet about this specific feature. WindowsForum notes that Microsoft’s general diagnostics docs say optional data can include app usage, but nothing explains why these particular tool slots exist.

It is also worth being clear that the follow-up coverage I found, including IT-Connect, is reporting on Li’s work rather than independently reproducing it. It is, currently, only one researcher’s well-documented teardown.

Why this matters to IT folks and homelabbers

While it seems dark, initially, a flag saying “this machine ran Python and Claude Code yesterday” is not going to leak your secrets. What bugs me is the why. Microsoft sells GitHub Copilot and competes directly with Anthropic, OpenAI, Google and Cursor for developers. An operating system quietly tallying which rival coding agents people run looks a lot like market research, gathered through the OS rather than through a product you chose to install.

For anyone managing Windows fleets, this is another reminder that “optional” telemetry is a moving target. Microsoft can light up new collection through server-side experiments without a visible update, so the only reliable control is the diagnostic data level itself. If your org has policies about developer tooling or data handling, it is worth confirming your endpoints are actually set the way you think they are.

What I am doing about it

The fix is the same boring one it has always been:

  1. On personal machines, go to Settings > Privacy & security > Diagnostics & feedback and turn off Send optional diagnostic data.
  2. On managed machines, set the diagnostic data level through Group Policy or Intune instead of trusting whatever was clicked during setup.
  3. When you build a new Windows VM in the lab, decline optional diagnostics during OOBE so you do not have to remember later.

The first one takes about ten seconds.

My take

I do not think Microsoft is reading anyone’s code here, and I appreciate that Li showed his receipts instead of hyping it. But “we only collect a bitmap” is not the reassurance it sounds like when the bitmap is a list of your competitors’ products, and the feature was turned on remotely with no announcement. Microsoft should explain what this is for. Until then, I would rather keep my optional data to myself.

What about you? Do you leave optional diagnostics on, or is this the nudge that gets you to flip the switch, or move that dev box to Linux?

Sources

Scroll to Top