A desktop monitor displays security settings and AI icons connecting to a cloud analytics dashboard.
Windows 11 can record whether certain coding agents and developer tools have run and queue that information for Microsoft through optional diagnostic reporting, according to a reverse-engineering investigation published by Xusheng Li on November 16, 2026. The finding is specific: a Windows component produces a compact usage bitmap, not a transcript of coding sessions. Li found no prompts or source code in that report, and his testing does not establish that the feature is enabled on every Windows 11 installation.

That distinction matters, but it does not make the privacy question disappear. Knowing which tools someone uses is information about their work—even when the report contains numbered categories rather than readable application names.

For readers, the immediate question is whether optional diagnostics are enabled. Microsoft documents a supported way to turn them off and says Windows remains equally secure and operates normally when sending only required diagnostic data. There is no need to begin by deleting system files or performing registry surgery.

A desktop monitor displays security settings and AI icons connecting to a cloud analytics dashboard. What Li found inside Windows​

The investigation centers on MissionControlAggregator.dll. Li initially used an AI assistant to identify potentially interesting functionality in Windows binaries, then followed executable-name comparisons through the component’s code and tested its behavior. The evidence goes beyond a list of suspicious-looking strings.

On his active test system, the Connected User Experiences and Telemetry service, known as DiagTrack, started AggregatorHost.exe, which loaded the DLL. The coding-agent checks were controlled by a feature named CLICodingAgentUsage, with feature ID 62580080.

The built-in matching lists cover two groups:

Matching groupExecutable names identified
Four command-line agent slotscopilot.exe, claude.exe, codex.exe, agy.exe
Five desktop-agent slotsgithub.exe, claude.exe, codex.exe, cursor.exe, antigravity.exe

The Windows Subsystem for Linux route checks corresponding names without the .exe extension. These are matching strings in the component, not proof that every matching executable is genuinely the product its name suggests.

Beyond agents, Li identified checks for node.exe, python.exe, dotnet.exe, gcc.exe, clang.exe, gdb.exe, javac.exe, devenv.exe and code.exe, among others. Those checks feed 12 broader developer-tool categories.

Some matches also require an expected publisher: Node.js Foundation for Node.js, Python Software Foundation for Python, and Microsoft Corporation for.NET and Visual Studio. Li found no additional publisher check in the compiler/debugger path containing GCC and Clang. That uneven validation becomes important when interpreting the tests.

It listens to events rather than scanning processes​

According to Li’s analysis, MissionControlAggregator does not independently walk through the list of running processes. It subscribes to existing Windows events and extracts application identities from them.

For native Windows applications, the inputs include:

  • KernelProcess.AppStateChangeSummary
  • Win32kTraceLogging.AppInteractivitySummary

For WSL, the relevant events are Microsoft.Windows.Lxss.Manager.Exec and ExecCritical, using the imageName field. A native Windows codex.exe can reach the command-line agent matcher through the app-state route; this is not a WSL-only finding.

A match updates a slot in the component’s state. A reporting routine named ReportCohorts<12,4,5> subsequently converts the recorded slots into bitsets and emits a self-describing TraceLogging event called Segments.SubCohorts. The numbers describe the 12 developer-tool categories, four command-line agents and five desktop agents.

Li gives the persistent-state location as:

C:\ProgramData\Microsoft\Diagnosis\AggregatorStorage\MissionControlAggregator$

The trailing dollar sign is part of the path shown in his original investigation.

Although the incoming events contain richer information—including runtime, focus time, interaction counters and resource usage—Li says this component appears to discard those details. Its report indicates which categories were observed during a collection window. The timestamps describe that window, not how many minutes someone spent using an agent.

The practical distinction: this is evidence of tool-usage classification, not evidence of the component reading a developer’s conversations or repository contents.

The renamed-executable test explains both the finding and its limits​

Li tested the command-line matching path by copying cmd.exe to a file named codex.exe and running it in an interactive Windows session. That was enough to set the Codex CLI slot. A separate run without an additional ETW capture also set the slot.

This is a useful controlled test: the executable’s name changed, but its underlying function did not. Windows Command Prompt had not suddenly become an AI coding assistant.

The emitted report contained CliAgentSubCohorts: 4, representing 1 << 2, the bit associated with the Codex CLI slot. A separate test using a copy named gcc.exe set developer-tool slot 3; its report contained SubCohorts: [8], representing 1 << 3. That category covers GCC, Clang and related compiler/debugger tools rather than identifying each separately.

There is also a timing caveat. Li describes reporting as following a roughly daily boundary. To obtain a report immediately, he moved the saved reporting-window start 25 hours into the past and restarted DiagTrack while tracing the aggregator. That manipulation accelerated the report; it should not be mistaken for an untouched observation of its normal reporting schedule.

The analytical takeaway is that an “observed Codex” bit is not conclusive proof of a genuine Codex session. The test strengthens the evidence that matching occurs while exposing a limitation in what the resulting measurement means.

What establishes the upload path?​

An Event Tracing for Windows event alone does not demonstrate transmission to Microsoft. Li therefore examined both the listener configuration and the diagnostic upload queue.

On the VM, the existing Diagtrack-Listener session subscribed to the input and output providers. Its KernelProcess keyword mask was 0x400000000002, including the summary keyword 0x400000000000. Its MissionControlAggregator mask was 0xe00000000000, including the report keyword 0x800000000000. Both subscriptions used level 255.

The more substantial evidence came from DiagTrack’s EventStore.db. Li inspected compressed JSON batches in its AsimovEvents table and found the agent report queued with:

  • Event name: Microsoft.Windows.MissionControlAggregator.Segments.SubCohorts
  • Privacy classification: isRequired: false
  • CLI-agent bitmap containing the observed Codex bit

That supports the conclusion that Windows queued the report for transmission as optional diagnostic data. The presented evidence is not a Microsoft server-side receipt, nor does it show this event being transmitted with optional diagnostics disabled. These are Li’s reported experiments, not independent WindowsForum laboratory results.

Microsoft’s broader documentation explains that optional diagnostics can include device activity and information about how Windows and applications are used. That provides context for the reporting category, but it does not independently confirm this particular event’s purpose.

Which Windows installations are affected?​

The active system was a Parallels Windows 11 ARM64 virtual machine running build 26200.9974, with MissionControlAggregator.dll version 10.0.26100.9818. Li says the feature was already enabled as a service experiment; he had not manually enabled it. DiagTrack was running with Automatic startup.

However, the public-ISO baseline VM had the coding-agent feature disabled by image default. That difference prevents a responsible conclusion that all Windows 11 PCs are doing this.

Separately, Li examined setup on a clean Windows 11 Home ARM64 installation configured for the United States. Optional diagnostic data was selected under the label “Include Optional,” and the completed installation had AllowTelemetry set to 3.

These observations concern two different controls: whether the agent feature is active and whether optional diagnostics are allowed. Neither observation proves a universal default across editions, architectures, regions or installation histories.

November 16 is the investigation’s publication date, not an established feature-launch date. The evidence does not identify a KB that introduced the behavior or provide a complete rollout map.

Microsoft also says some optional diagnostic collection is sampled rather than collected from every eligible device. That general policy reinforces the need for caution when extrapolating from one machine, although it does not establish the sampling rules for MissionControlAggregator.

How to turn off optional diagnostic reporting​

For a personal Windows 11 PC, use the supported privacy control:

  1. Open Settings.
  2. Select Privacy & security.
  3. Open Diagnostics & feedback.
  4. Set Send optional diagnostic data to Off.

The immediate success condition is that the switch shows Off. During setup, users can also decline optional diagnostics rather than accept their inclusion. Microsoft documents the diagnostic controls, and Li identifies this setting as relevant to the optional report he observed.

Turning off optional diagnostics is not the same as disabling every diagnostic function. Required diagnostic data remains a separate category, and Microsoft explains that selecting optional diagnostics includes required data as well. Treat this as a control over the additional reporting level—not a universal “send nothing” switch.

There are two important exceptions to a straightforward personal-device walkthrough:

  • A work or school PC may be policy-managed. Microsoft’s Windows Privacy Compliance Guide says administrators can restrict users’ ability to change the diagnostic setting. If the control is unavailable, ask IT to verify the effective policy rather than attempting to bypass it.
  • Windows Insider Preview has additional requirements. Microsoft’s Insider documentation says optional diagnostic data must be enabled to run Insider Preview builds. Preview participants should account for that requirement before changing the setting.

Disabling the optional upload level should not be described as proof that all local application events stop being generated or that every saved aggregator record disappears. Li’s investigation distinguishes local events, persistent state and the upload queue; changing one reporting control is not evidence that all three have been erased.

How to inspect without overinterpreting the results​

Administrators investigating a particular machine should distinguish component presence, feature activation and queued reporting. They are not interchangeable findings.

Li’s investigation includes administrative PowerShell checks such as:

Code:
Get-Service DiagTrack | Format-Table Name,Status,StartType -AutoSize
tasklist /m MissionControlAggregator.dll

The first shows the service’s state and startup type. The second can identify a process with the DLL loaded. Neither command, by itself, proves that the coding-agent feature is enabled or that a particular usage event has entered the upload queue. Li’s stronger conclusion depended on matching tests, saved state, emitted events and queue inspection together.

For a supported view of diagnostic reporting, Microsoft offers Diagnostic Data Viewer, documented for Windows 11 version 21H2 and later. It requires installation from Microsoft Store and enabling diagnostic-data viewing in Settings. Microsoft describes it as a way to review diagnostic information the device is sending.

The viewer should not be presented as a guaranteed one-click detector for this specific experiment. Its useful role is broader: making diagnostic reporting inspectable without asking ordinary users to modify reporting-window timestamps or unpack internal databases.

Also avoid confusing the relevant switch with App diagnostics. Microsoft documents that separate privacy page as controlling applications’ access to diagnostic information about other apps. The optional Windows reporting control discussed here is under Diagnostics & feedback.

Why a small bitmap still deserves scrutiny​

The absence of prompts and source code substantially narrows this finding. It does not make tool-choice data meaningless. A bitmap whose categories are known can still communicate whether a machine was classified as using particular agents or development environments.

Microsoft’s published rationale for Windows diagnostics includes reliability, troubleshooting and product improvement. Its data-collection summary explicitly places product and service usage within the optional diagnostic framework. That is the company’s general explanation—not a demonstrated explanation for why these particular coding-agent slots exist.

The unresolved questions are therefore specific: how widely the feature is enabled, when it was introduced, why these tools were selected and how Microsoft uses this particular report. The investigation does not establish answers, and the executable list alone cannot supply them.

The appropriate response is neither “Windows is reading all your code” nor “it is only a few bits, so nothing matters.” Li documented a targeted usage-reporting mechanism with meaningful limits. For developers and IT administrators, the actionable lesson is to review optional diagnostic settings deliberately—and judge telemetry by the information it conveys, not merely the size of its payload.

 

References

  1. Windows 11 Collects Data on Coding Agents and Developer Tools You Use - news.lavx.hu news.lavx.hu 2026-11-17T00:48:06.359000+00:00
  2. Windows 11 Is Collecting Data on Which Coding Agents and Developer Tools You Use :: Xusheng Li xusheng.dev
  3. Configure Windows diagnostic data in your organization - Windows Privacy | Microsoft Learn learn.microsoft.com