Windows 11 has a component that watches for coding agents and developer tools, records which ones you have run, and, unless you have turned off optional diagnostic data, uploads that record to Microsoft. The finding comes from a reverse-engineering writeup by xusheng.dev, who pulled the relevant strings out of a system DLL called MissionControlAggregator.dll and then confirmed the behavior on a live ARM64 virtual machine.

The list embedded in the DLL is blunt about what it is looking for. Four CLI-agent slots check for copilot.exe, claude.exe, codex.exe, and agy.exe, with extensionless equivalents on the WSL route. Five desktop-agent slots check for github.exe, claude.exe, codex.exe, cursor.exe, and antigravity.exe. The feature guarding those checks has a name that leaves little to interpretation: CLICodingAgentUsage, feature ID 62580080.

This is not a rumor about a future build. The author found the feature already enabled as a service experiment on a Parallels Windows 11 ARM64 VM running build 26200.9974, with MissionControlAggregator.dll at version 10.0.26100.9818. He had not enabled it himself.

What the mechanism actually does

The component does not scan running processes. It subscribes to Windows events that already exist: KernelProcess.AppStateChangeSummary and Win32kTraceLogging.AppInteractivitySummary for native programs, and Microsoft.Windows.Lxss.Manager.Exec for WSL. A regular Windows codex.exe reaches the CLI-agent checks through the app-state summary route, so this is not limited to Linux binaries inside WSL.

A match sets a slot in an internal state. A function called ReportCohorts<12,4,5> later turns those slots into a bitset and emits a TraceLogging event named Segments.SubCohorts. The numbers map to twelve broader tool categories, four CLI agents, and five desktop agents. The report carries a bitmap, not a log. The source events include runtime, time in focus, and resource usage, but this component discards those details and keeps only which tools were observed.

The author tested it by copying cmd.exe to a file named codex.exe and running it. The name alone was enough. The saved state at C:\ProgramData\Microsoft\Diagnosis\AggregatorStorage\MissionControlAggregator$ recorded the value 2 in CLI slot 2, the Codex slot. A second test with a copy named gcc.exe set developer-tool slot 3, the compiler/debugger category shared by GCC and Clang. The emitted event read CliAgentSubCohorts: 4, which is 1 << 2, the Codex bit.

The part that matters most is what happens next. The author decompressed the JSON batches in DiagTrack’s EventStore.db and found the agent-usage report sitting in the upload queue, tagged "privacy": { "isRequired": false }. That tag marks it as optional diagnostic data. On a clean Windows 11 Home ARM64 install from the latest official ISO with the region set to the United States, the Diagnostic data toggle was already on and labeled Include Optional, and AllowTelemetry was 3. Optional diagnostics are enabled by default in that setup.

Why a newsroom should care

The obvious read is a privacy story, and it is one. A bitmap of which AI coding agents a developer runs is a weak signal in isolation. Aggregated across a fleet, it becomes a market-research dataset that no vendor disclosed collecting. Microsoft can see, at minimum, the relative uptake of Codex, Claude, Copilot, Cursor, and Antigravity among Windows developers who left defaults alone. That is competitive intelligence about the AI tools market, gathered as a side effect of an operating system.

The narrower read is more interesting for anyone building AI software. The list is a census of the agentic coding market as Microsoft’s diagnostics team sees it. Four CLI agents and five desktop agents, each with a dedicated bit. The presence of antigravity.exe and agy.exe alongside the better-known names is a reminder that the agent-tooling category is crowded and moving fast enough that an OS vendor felt the need to enumerate it. Microsoft is not just watching its own Copilot. It is watching the whole field.

There is a real argument that this is unremarkable. Windows has collected telemetry on installed applications for years, and the author is careful to note that the public ISO’s baseline VM had the coding-agent feature disabled by image default. The evidence does not support the claim that this is active on every Windows installation. That distinction matters, and anyone repeating the finding should keep it.

But the design choice is still worth naming. The component assigns a bit per agent rather than a bit per category, which is more granularity than a diagnostic system needs to fix a crash. It checks the publisher for Node.js, Python, and .NET but not for the compiler group, and it does not appear to check publishers for the agent slots at all. A renamed cmd.exe was enough to trigger detection. That is a fingerprinting surface, not a stability metric.

What builders should take from it

If you ship a coding agent, your users’ Windows machines may now report that they run it. There is no API, no notification, and no per-app consent flow. The only switch is the global optional-diagnostics toggle under Settings, Privacy and security, Diagnostics and feedback. Enterprises that manage fleets through policy can turn it off centrally. Individual developers who clicked through setup in the US almost certainly did not.

The interesting question is not whether Microsoft should collect this. It is whether the AI tooling market has any shared expectation about it. Model providers publish usage policies, rate limits, and data-retention terms. None of them, as far as we can tell from this writeup, anticipated that the operating system underneath would independently count which agents get launched. That is a gap in the ecosystem’s threat model, and it is now documented in a single blog post with PowerShell output.

Microsoft has not commented on the finding. The author’s reproduction steps are specific enough that anyone with a Windows VM can check them: look for DiagTrack running as Automatic, find AggregatorHost.exe as its child, and inspect MissionControlAggregator.dll inside it. The bit for Codex is 1 << 2. The bit for the compiler group is 1 << 3. The report is called Segments.SubCohorts, and it is queued for upload unless someone turns the toggle off.

What to watch is whether Microsoft changes the default, adds a per-app disclosure, or says nothing and lets the finding fade. The precedent set here will shape how the next wave of OS-level AI tooling telemetry gets framed. For now, the census runs quietly, one bit per agent, on machines where nobody opted in.