← Back to all posts
News

Windows 11 Logs Which Coding Agents You Run and Queues It for Upload to Microsoft

November 17, 2026 · 02:24 UTC · News
On this page 7 sections
Windows 11 Logs Which Coding Agents You Run and Queues It for Upload to Microsoft

TL;DR

Reverse engineering by Xusheng Li shows Windows 11 ships a diagnostics component that keeps a name-based watchlist of AI coding agents. MissionControlAggregator.dll, loaded inside a process started by the DiagTrack telemetry service, matches executables including claude.exe, codex.exe, copilot.exe and cursor.exe, records which of them were seen on your machine, and emits a report roughly once a day. The report is queued for upload to Microsoft as optional diagnostic data, and in the US clean install Li tested, that setting is on unless you turn it off. No prompt text and no source code are collected; the payload is which tools were observed, one bit each.


At a glance
What the DLL records
9
agent slots, 4 cli plus 5 desktop
12
broader developer tool categories
62580080
feature id guarding the agent checks
0
prompts or source code in the payload
~1/day
cadence of the usage report
Nov 16
date the analysis went up

The watchlist is explicit

Li found the names while hunting undocumented functionality in the latest Windows binaries, then followed the references in Binary Ninja. Four slots watch CLI agents: copilot.exe (Copilot), claude.exe (Claude), codex.exe (Codex) and agy.exe, with extensionless equivalents checked on the WSL route. Five more watch desktop agents: github.exe, claude.exe, codex.exe, cursor.exe (Cursor) and antigravity.exe (Antigravity).

Ordinary developer tools get the same treatment. node.exe, python.exe, dotnet.exe, gcc.exe, clang.exe, gdb.exe, javac.exe, devenv.exe and code.exe feed the twelve broader tool categories. Some slots also verify the publisher: Node.js Foundation for node.exe, Python Software Foundation for python.exe, Microsoft Corporation for the .NET and Visual Studio slots. The GCC and Clang group gets no publisher check in this path.

The watchlist
Agent names compiled into the DLL
copilot.execlaude.execodex.exeagy.exegithub.execursor.exeantigravity.exe
claude.exe and codex.exe each appear in both the CLI and the desktop lists. Source: Xusheng Li's binary analysis.

From sighting to upload queue

The component does not scan processes itself. It subscribes to ETW events Windows already emits: KernelProcess.AppStateChangeSummary and Win32kTraceLogging.AppInteractivitySummary for Windows apps, and Microsoft.Windows.Lxss.Manager.Exec for WSL, keyed on the image name. A match sets a slot in a state file under C:\ProgramData\Microsoft\Diagnosis\AggregatorStorage, and on a timer the aggregator packs the slots into bitsets and emits a Segments.SubCohorts event: twelve tool categories, four CLI agents, five desktop agents.

How it works
From process name to Microsoft upload queue
Observe
ETW app events
windows app summaries, wsl exec included
Match
Nine agent names
name-only check in this path
Store
One bit per tool
saved under programdata, diagnosis
Report
Segments.SubCohorts
emitted on a roughly daily boundary
Upload
DiagTrack queues it
tagged isrequired: false, optional data
The report carries which tools were seen, not how long they ran or what was typed. Source: Xusheng Li's analysis.

The report is a bitmap, which keeps it small and vague at the same time: picture a hotel chart of which amenities you touched rather than your itinerary. Runtime, time in focus and interaction counts exist in the source events, and this component throws them away. The timestamps describe the collection window, not your usage.

The test that makes it hard to dismiss

Strings in a DLL show intent, and Li went further and made the chain run. On a Windows 11 ARM64 VM, build 26200.9974 with DLL version 10.0.26100.9818, the CLICodingAgentUsage feature was already enabled as a service experiment nobody had asked for, and the DiagTrack service was running with startup type Automatic, hosting an AggregatorHost.exe process with the DLL loaded.

The check matched on file name alone. Li copied cmd.exe to a file named codex.exe and ran it, and the aggregator set CLI slot 2, the Codex bit, storing the value 4. A copy named gcc.exe set developer-tool slot 3. Windows logged a frontier coding agent on the strength of a filename, no questions asked.

The upload half holds up too. With optional diagnostic data allowed, Li found the Segments.SubCohorts event sitting in DiagTrack's upload queue inside its EventStore database, tagged isRequired: false. On a clean Windows 11 Home install from the latest official ISO, region set to the United States, the Diagnostic data toggle came up enabled and labeled Include Optional, with AllowTelemetry at 3, before any privacy settings were touched.

The proof
A renamed command prompt set the Codex bit
slot 2
codex bit set by a renamed cmd.exe
26200.9974
windows 11 build tested, arm64 vm
false
isrequired tag on the queued report
3
allowtelemetry on the us clean install
Source: Xusheng Li's analysis, Nov 16, 2026.

The caveats matter here

Two limits are worth knowing before you rip out your SSD. The public ISO's baseline VM had the coding-agent feature disabled by image default, so the evidence does not say every Windows 11 install is reporting agent use today; service flighting can switch it on later, which is what Li's VM demonstrated. And this is one researcher's analysis, published Nov 16 with full method, commands and event names, with no Microsoft comment and no independent replication yet; the Hacker News thread it drew, 244 points at time of writing, contains arguments about Linux but no technical refutation.

Microsoft's own documentation already discloses the bucket. Optional diagnostic data "may also include information about the websites you browse, device activity (sometimes referred to as usage), and enhanced error reporting", the specific items collected are "subject to change", and the device "will be just as secure and will operate normally" on required data alone. App usage collection was already public; the coding-agent roster inside the machinery is the new detail.

What to do about it

The off switch is Settings, then Privacy and security, then Diagnostics and feedback, then Send optional diagnostic data. You can decline during setup instead, and the Diagnostic Data Viewer shows what a running device actually transmits. On managed fleets, the diagnostic data level is set by organizational policy, so this is a conversation with your IT department rather than a settings toggle.

Two builder-specific angles. The WSL route means running the Linux build of an agent does not dodge the matchers. And if you work under an agreement that treats tool choice as confidential, your agent stack now inventories itself and leaves the machine, whatever telemetry the agents themselves have disabled. Renaming your binaries defeats the name check, which is either comforting or damning, depending on how much you trusted the data in the first place.

Key Takeaways

  • Windows 11's MissionControlAggregator.dll contains a feature named CLICodingAgentUsage, feature ID 62580080, with slots for nine coding agents: copilot.exe, claude.exe, codex.exe and agy.exe on the CLI side, github.exe, claude.exe, codex.exe, cursor.exe and antigravity.exe on the desktop side.
  • It subscribes to existing ETW events including WSL exec, sets one bit per tool observed, persists the state under ProgramData, and emits a Segments.SubCohorts report on a roughly daily cadence. No prompts or source code are collected.
  • Li proved the chain end to end: a cmd.exe renamed to codex.exe set the Codex bit, and with optional diagnostic data allowed the report appeared in DiagTrack's upload queue tagged isRequired: false.
  • In a US clean install of Windows 11 Home from the latest ISO, optional diagnostic data was on by default with AllowTelemetry at 3. The off switch lives under Settings, Privacy and security, Diagnostics and feedback.
  • Caveats: the public ISO baseline had the feature disabled by image default, so it is not confirmed active on every install, and this is one researcher's analysis with no Microsoft comment and no independent replication yet.
  • Microsoft's documentation already lists app activity and device usage under optional diagnostic data and says collected items are subject to change, so the disclosure bucket existed; the agent-specific roster is the new fact.

Sources: Xusheng Li's reverse engineering analysis, initial finding on X, Microsoft diagnostics and privacy documentation, Hacker News discussion

AIWindows 11TelemetryPrivacyCoding AgentsMicrosoftReverse EngineeringDiagnostic Data
CONSOLE
$