Back to Explore
NewsVibeFix editorialUpdated Nov 20, 2026

Windows 11 is logging which coding agents you run — on by default, but you can turn it off

On 16 November 2026 researcher Xusheng Li found Windows 11's MissionControlAggregator records which coding agents you run — Codex, Claude, Cursor and more — as a bitmap sent with optional diagnostic data. A clean US install enables it by default, the public ISO baseline does not, and Microsoft has not commented.

A laptop on a dark desk showing a Windows 11 desktop with a code editor open

In short: the coding agents you run may be recorded by Windows 11

On 16 November 2026, independent researcher Xusheng Li published "Windows 11 Is Collecting Data on Which Coding Agents and Developer Tools You Use" on his personal blog, disclosing that a Windows 11 system component records which AI coding agents and developer tools have run on a machine, and queues a summary for upload to Microsoft when "optional diagnostic data" is enabled. The post reached the Hacker News front page the same day. It was read in full (268 lines); all core facts below come from it. As of 17 November, Microsoft had not commented.

The discovery itself is worth telling. Li wrote that he asked an AI to help him look through the latest Windows binaries for undocumented features. One list immediately caught his attention: executable names for coding agents such as codex.exe, claude.exe, copilot.exe, and cursor.exe, alongside familiar developer tools such as python.exe, dotnet.exe, gcc.exe, and code.exe. Why would a Windows system DLL need a list of coding agents and developer tools? Following the references led from a handful of string comparisons to a running Windows component that records which agents and tools have been used, saves that information locally, and reports it to Windows Diagnostics. Unless you have opted out of optional diagnostic data — which is enabled by default, at least in the US setup flow he tested — the data is uploaded to Microsoft.

This matters directly to vibe-coding readers. Claude Code, Codex, Cursor, and Copilot are the tools we use every day, and a typical vibe coder's machine has several of them installed at once. If the operating system quietly ticks off which of these tools you run, then "which tools you code with" itself becomes a telemetry signal. This is not "Microsoft reading your code" — the original post is explicit that the report contains no prompt text and no source code — but "knowing what you use" is still something to take seriously. The rest of this article lays out the finding, the evidence, and the limits, one by one.

What it actually records: a bitmap, not a log

The most important distinction first: the component reports a bitmap, not a log. The names live in MissionControlAggregator.dll. Reverse engineering shows four CLI-agent slots checking copilot.exe, claude.exe, codex.exe, and agy.exe, with extensionless equivalents on the WSL route, and five desktop-agent slots checking github.exe, claude.exe, codex.exe, cursor.exe, and antigravity.exe. Beyond agents, the component checks for node.exe, python.exe, dotnet.exe, gcc.exe, clang.exe, gdb.exe, javac.exe, devenv.exe, and code.exe, among others, feeding twelve broader developer-tool categories. Some checks also verify the publisher: the Node.js slot expects "Node.js Foundation", Python expects "Python Software Foundation", and the .NET and Visual Studio slots expect "Microsoft Corporation", while the compiler/debugger group containing GCC and Clang has no additional publisher check in this path.

A successful match sets a slot in the aggregator's persisted local state. Roughly once a day, a timer fires ReportCohorts<12,4,5>, turning those slots into bitsets and emitting a report named Segments.SubCohorts. The numbers 12, 4, and 5 correspond to twelve tool categories, four CLI agents, and five desktop agents. Rather than writing executable names into every report, the component assigns each category a bit: a set Codex bit means the matching tool was observed during the reporting window. The report's timestamps describe the collection window, not how long each agent was used. The source events also include runtime, time in focus, interaction counters, and resource usage, but this component appears to discard those details and record only a bitmap of which agents were observed.

That design determines the privacy meaning. A bitmap is genuinely "lighter" than a detailed log: it does not know what code you wrote, what you asked, or how long you worked. But "lighter" is not "meaningless". Knowing whether someone uses Cursor or Codex, Python or .NET, is enough to sketch a technical profile — especially combined with other diagnostic data Microsoft already holds. Li's own judgement is restrained and accurate: although he found no prompt text or source code in this report, collecting which agents and developer tools someone uses already has significant privacy implications. Do not inflate it into keylogging, and do not dismiss it as "just statistics".

How it knows: no process scanning, just existing event streams

The next question is whether the component scans your running processes. Li's answer is no. Following Aggregator::OnLoad showed that it subscribes to existing Windows events instead. For Windows programs, it consumes KernelProcess.AppStateChangeSummary and Win32kTraceLogging.AppInteractivitySummary, which provide application identities that its handlers parse and pass to the matchers. For WSL, it consumes Microsoft.Windows.Lxss.Manager.Exec and ExecCritical, using the imageName field. A regular Windows codex.exe can therefore reach the CLI-agent checks through the app-state summary route; this is not limited to Linux programs running inside WSL.

The flow looks like this: Aggregator::OnLoad subscribes to the three event sources, feeding CLI-agent matchers or desktop-agent matchers; Aggregator::OnRelativeTimerExpire calls MissionControlAggregatorLogging::ReportCohorts, emitting the Segments.SubCohorts report; and Aggregator::OnUnload or an intermediate timer saves the persistent state to disk. The structure says something important about the nature of this: the component is more like a "freeloader" summariser than a standalone monitor. It produces no raw signal of its own; it adds a layer of "do I recognise this name?" on top of event streams Windows already produces. Microsoft did not install a new camera for this; it added a new classifier to an existing diagnostic pipeline.

The component runs under DiagTrack, the Connected User Experiences and Telemetry service. DiagTrack starts a separate AggregatorHost.exe process, and that process loads the DLL. Li verified this on his Parallels Windows 11 ARM64 virtual machine, build 26200.9974: the DLL version was 10.0.26100.9818, DiagTrack was running with startup type Automatic, and AggregatorHost.exe was its child process, with the module listing placing MissionControlAggregator.dll inside that child. The process IDs line up — DiagTrack in process 10148, AggregatorHost.exe as process 1500. Strings and code were a lead; a running process is evidence, and this step moves the finding from "the DLL contains these names" to "the component is actually running".

The experiment: rename cmd.exe to codex.exe and it triggers

The most convincing part of the post is the live test. Li copied cmd.exe to a file named codex.exe and ran it in an interactive Windows session. The name alone was enough for the component to recognise it as Codex. He captured an AppStateChangeSummary for the running process from the KernelProcess provider, with LaunchCount=1 and TerminateCount=0. The aggregator's saved state contained the value 2 in CLI slot 2, the Codex slot. A separate run without an additional ETW trace also set that slot, confirming that detection worked through Windows' existing diagnostic collection.

The state is persisted at C:\ProgramData\Microsoft\Diagnosis\AggregatorStorage\MissionControlAggregator$. The report normally follows a roughly daily boundary. To capture it immediately, Li temporarily moved the saved reporting-window start 25 hours into the past and restarted DiagTrack with a trace listening for the aggregator. The resulting Segments.SubCohorts event read: provider Microsoft.Windows.MissionControlAggregator, time 2026-11-16T12:34:29.125116Z, version 0.1.5, CliAgentSubCohorts 4, DesktopAgentSubCohorts 0, FlushedFromSave true. The value 4 is 1 << 2: the Codex CLI bit. The name had made it through the input summary, the matcher, the persisted state, and the emitted report — the chain is complete.

He also tested the developer-tool path separately, using a copy of cmd.exe named gcc.exe. It set developer-tool slot 3. With optional diagnostic data allowed, he found the resulting Segments.SubCohorts event in DiagTrack's upload queue with SubCohorts: [8] (1 << 3). This is the compiler/debugger category shared by GCC, Clang, and related tools, rather than a separate bit for each executable. That detail is telling: the twelve tool categories are category-level bits, while the four CLI agents and five desktop agents get product-level bits. The granularity Microsoft wants is product-precise for agents and category-precise for tools. That asymmetry itself shows where the new emphasis lies: the point is not developer tools, it is agents.

Where the data goes: found in the upload queue

An ETW event by itself would not establish that anything leaves the machine — ETW is used for plenty of local diagnostics. The more revealing question was who was listening to this particular report. The VM already had a Diagtrack-Listener session subscribed to both ends of the path: the KernelProcess provider and the Microsoft.Windows.MissionControlAggregator provider, both with MatchAnyKeyword values covering the report's keywords, at level 255. The subscription showed where the report could go. To check whether Windows actually queued it for transmission, Li inspected DiagTrack's EventStore.db and decompressed the JSON batches in its AsimovEvents table. That revealed an important condition: the agent-usage report is treated as optional diagnostic data.

When optional diagnostic data is allowed, Windows queues the coding-agent usage report for upload to Microsoft. Li verified this by finding the agent event itself in DiagTrack's upload queue: name "Microsoft.Windows.MissionControlAggregator.Segments.SubCohorts", privacy isRequired false, data Version 0.1.5, CliAgentSubCohorts [4], DesktopAgentSubCohorts [0], FlushedFromSave true. The report was not only generated but queued for upload. That closes the chain: executable name, matcher, persisted state, ETW report, upload queue — each step backed by a live test.

That optional diagnostic data is enabled by default is the point most ordinary users should notice. Li checked with a clean installation of Windows 11 Home ARM64 using the latest official ISO, with the region set to the United States. Before changing any privacy settings, the Diagnostic data toggle was already enabled and labelled "Include Optional". After completing setup with those defaults, AllowTelemetry was 3, confirming that optional diagnostic data was enabled. His wording is careful: optional diagnostic data is enabled by default, at least in the US setup he tested. If you did not opt out during installation or disable it later, it is still enabled and your coding-agent usage data is being uploaded. You can decline optional diagnostics during setup or turn off "Send optional diagnostic data" later under Settings, Privacy & security, Diagnostics & feedback. Microsoft's own diagnostics and privacy guide explains how to change it, and the original post links to it.

Do not panic: the scope is narrower than the headline, but bigger than a false alarm

Headlines easily suggest "every Windows 11 machine reports the agents you use". The evidence in the post does not support that, and Li draws the line himself. On his Parallels ARM64 VM, build 26200.9974, with DLL version 10.0.26100.9818, CLICodingAgentUsage (feature ID 62580080) was already enabled as a service experiment, and he had not manually enabled the feature. But the public ISO's baseline VM had the coding-agent feature disabled by image default. So the evidence does not support saying it is active on every Windows installation. What can be said is that a staged rollout is underway, DiagTrack ships enabled and set to automatic, and the report flows out whenever optional diagnostic data is on.

As of early 17 November UTC there was no Microsoft statement about the feature. Windows' own admin documentation describes optional diagnostic data as including app activity such as which programs launch on a device — consistent with this finding, but it never names coding agents. So the situation is subtle: nothing here contradicts Microsoft's published description of optional diagnostic data in general, but nothing tells users plainly "we will record which coding agents you use" either. The "Include Optional" choice made during setup actually covers a collection whose specifics users do not know. That is why it is controversial: the issue is not how much data there is, but whether consent is real.

Two common misreadings need correcting. First, this is not "Microsoft watching your code". The report contains no prompt text, no source code, no file names, and no usage duration — only a set of "observed during this window" bits. Second, this is not an isolated "gotcha" but a feature in staged rollout: the same report contains both an experiment environment where it is enabled and a baseline image where it is disabled by default. Its true state is "rolling out", not "already on every machine". Keeping both boundaries clear avoids understating it and overstating it.

What this means for vibe coders

The practical step first: check your settings. If you write code on Windows 11, spend a minute in Settings, Privacy & security, Diagnostics & feedback, and see whether "Send optional diagnostic data" is on. Turn it off and this agent-usage report will not enter the upload queue. The same applies when installing a new system — do not click through the diagnostic-data step on defaults. Company-issued machines deserve an extra caution: the device may be managed by administrators through policy, the toggle you see may not be the final word, and compliance requirements may require diagnostic data to stay on. Ask IT first rather than quietly editing the registry yourself.

One easily missed point: WSL users are in scope too. The component consumes WSL execution events through Microsoft.Windows.Lxss.Manager.Exec, matching on the imageName field. Running Codex CLI inside WSL is recorded by the CLI slot just the same. Conversely, matching recognises only executable names, not behaviour: Li's experiment proved that renaming cmd.exe triggers the Codex slot. That name-only matching cuts both ways. It means false positives are easy — any same-named file counts. It also means the judgement is very shallow: the component does not know, and does not care, whether what you ran is really Codex. The shallow side has an upside — it genuinely cannot reach your code — and a downside: it counts the appearance of a name, not a real use.

The deeper question is about the visibility of "tool choice" itself. Until now, switching tools was a private matter: moving from Cursor to Codex, or keeping four agents installed and rotating between them, was known only to you. Now there is an observer, and the observer is the operating system. A single bitmap is light, but collected daily over months and combined with other diagnostic data Microsoft already holds, a person's tool preferences, switching rhythm, and even working habits become inferable. For individual developers this is mainly a question of informed consent; for companies there is an added compliance layer: the list of third-party agents employees use on company machines is flowing to a vendor as diagnostic data. The list contains no code, but it is still intelligence about "what tools this company uses to do its work".

Judging the finding: solid method, single sample

A piece of independent reverse engineering should be judged on method and sample together. On method, this one is solid: start from binary strings, follow references to the matchers, confirm the event subscriptions, verify the detection logic with a rename test, and finally find the report itself in the upload queue — each step with reproducible commands and outputs. Worth noting: Li is open that the initial lead came from asking an AI to look through the binaries. That does not weaken the conclusion — the AI supplied leads, while the verification was his own hands-on rename test and queue inspection — but it reminds readers that the "discovery" part had AI involvement, and what really holds up is the test chain that follows, not the initial list.

On sample, the post is singular: one Parallels ARM64 virtual machine, one build version, one clean US install. How far rollout has progressed on x86 machines, whether install defaults differ in other regions, and whether Enterprise and Education editions behave differently are all unanswered. Discussion on Hacker News concentrated on exactly that: some readers confirmed the DLL's presence, others cautioned that rollout state may differ by channel. Microsoft's silence leaves those questions open for now. Until Microsoft responds or more people reproduce the work, the most accurate statement is this: Windows 11 is in staged rollout of an optional-diagnostic feature recording coding-agent usage; a clean US install enables optional diagnostic data by default; turning it off stops the upload — not "your computer is being monitored".

Finally, the limits of this article. All core facts come from Xusheng Li's personal blog post of 16 November 2026, read in full; the Hacker News front-page position was confirmed by a front-page query on 17 November. Microsoft's lack of response is as of before publication on 17 November. Every figure — four CLI slots, five desktop slots, twelve tool categories, feature ID 62580080, build 26200.9974, DLL version 10.0.26100.9818, AllowTelemetry value 3 — comes from the original post's own tests, cited with the original's qualifications intact. If you do vibe coding on Windows 11, go and look at your diagnostic settings now: deciding with full information whether to leave it on or turn it off is more useful than arguing about whether Microsoft is watching you.

Windows 11 coding-agent telemetry chain diagram: executable-name matching to bitmap report to the optional diagnostic data upload queue, with the staged-rollout limit noted

Views 5Comments 0

Comments (0)

ME
0/1000
Loading comments...

Sources

Browse projectsPublish your project

Related articles

A developer with glasses types on a laptop at a wooden desk, an external monitor to the left showing C code on a dark background, a small plant on the desk
News
Microsoft's Coding Model Stops Playing Second Fiddle: MAI Code 2 Scores 96% of Opus 5.5 for 5% of the Cost — Read the Three Caveats First

On 17 November 2026, Microsoft launched MAI Code 2 and Code 2 Flash: the larger model has 77 billion active parameters and a one-million-token context, averaging 63.2% across four benchmarks — close to GPT-6.1 Sol and Claude Opus 5.5 — at $0.309 per task. Read the three caveats first: promotional pricing, the lowest score of the three, and no long-horizon work.

Developer WorkflowProduct LaunchModel Updates
A developer types at a dual-monitor workstation; the left screen shows a dark code editor, the right a testing dashboard with green checkmarks and bar charts
News
Did Your Agent Get Better or Worse? AWS Open-Sources UAEF to Measure Agents from Every Framework with One Ruler

On 11 November 2026, AWS open-sourced UAEF: adapters turn traces from LangGraph, LangChain, Strands and Bedrock AgentCore into one format, then 36 metrics across 7 dimensions score them, flagging regressions past 5%. The judge model is hardened too — scores must cite verbatim evidence, and never decide safety.

Developer WorkflowProduct LaunchTesting & Quality