Four command-line slots. Five desktop slots. One bit apiece. That is the report Xusheng Li found inside MissionControlAggregator.dll, a Windows 11 component that notes which coding agents you run — Copilot, Claude, Codex and Antigravity at the command line; GitHub, Claude, Codex, Cursor and Antigravity on the desktop — and, unless you declined "optional" diagnostic data, queues the result for upload to Microsoft. On the clean US install Li tested, that option was already on.

Most of the 500-comment Hacker News thread argued about privacy. The more telling detail is how the report is built. The same component sorts a dozen categories of ordinary developer tools, and GCC and Clang share a single compiler-and-debugger bit. Coding agents get a bit per vendor. Windows counts compilers by category and agents by brand, and only one of those is a question about Windows.

The evidence is in what the component keeps and what it throws away. Its inputs are events Windows already collects about processes and app interactivity, and they carry far more than a name:

The source events also include runtime, time in focus, interaction counters, and resource usage. This component appears to discard those details and record only a bitmap of which agents were observed.
Xusheng Li

A diagnostic would care whether codex.exe crashed, hung or pinned a core. This one doesn't check whether it is Codex at all: Li copied cmd.exe to a file named codex.exe, ran it, and the Codex bit lit up. The matcher reads the name on the box. The report it feeds is called Segments.SubCohorts — the vocabulary of a growth team, not a reliability team — and the flag that gates it, CLICodingAgentUsage, was already enabled as a service experiment on Li's own VM, where he hadn't turned it on.

Put together, the report answers one question across the Windows installed base: of the developers using coding agents, how many use Microsoft's and how many use someone else's. That is a market-share survey of the most contested category in software, run by a company with a product in the race, on its rivals' customers' machines.

Windows counts compilers by category and agents by brand, and only one of those is a question about Windows.

The best defense is that the data is small, and it is. No prompts, no code, no usage time; Li found none of those in the report. Microsoft also has a real reason to know how many of its users run agents, since WSL, Windows Terminal and Dev Drive are built around what developers actually run. Both points are true, and the report's own design answers them. A category bit — a coding agent ran today — would tell Windows everything it needs to plan for agent workloads. It is exactly what Microsoft settled for with compilers. Splitting that bit by vendor adds nothing for Windows and a great deal for Copilot.

Data minimization protects the person, not the market. One developer's handful of bits is harmless. Multiplied across Windows, those bits are the adoption curves of Anthropic, OpenAI, Anysphere and Google, held by the company selling against all four.

The honest version of this report already exists, in the same DLL. It counts GCC and Clang as one. Claude and Copilot each get their own.