Originally published at https://blog.joonquixote.com/en/posts/wangs-agent-still-shows-up/.
At a game company in China, an employee quit, and an AI agent trained on his work records took over his job.
It sounds like a curiosity from a distant news cycle.
Except that around the same time, Meta was drawing up plans to capture employees’ keystrokes and screens to train AI agents of its own.
This post traces both stories and sorts the trend into two forces: companies harvesting how people work from the inside, and professionals selling how they work from the outside.
Then it turns to the part that hits closest to home for developers: the session logs that coding agents like Claude Code and Codex quietly leave on your machine.
By the end, you’ll have a clearer way to think about the work traces you produce every day, and one question worth asking before you need the answer.
Yesterday’s coworker, today’s agent
In April, the South China Morning Post reported on a game company in Shandong province that turned a former employee into an AI-powered digital worker.
The company trained the system on documents and work patterns the departed HR specialist left behind, reportedly with his consent, and put it to work on the job he used to do.
The agent introduces itself in the company chat as the former employee’s avatar, then fields questions, schedules meetings, and produces slide decks and spreadsheets.
The backlash was immediate: there is something deeply unsettling about your replica continuing to work after you’ve left the building.
Around the same time, a GitHub project called colleague.skill went viral on Chinese social media.
Feed it someone’s chat history, emails, code, and documents, and it produces an agent that mimics their voice and their way of making decisions.
The pitch promised to “turn cold goodbyes into warm skills” and welcomed users to cyber-immortality.
It started as a tongue-in-cheek stunt, but it struck a nerve, and copycats followed: boss.skill, mentor.skill, and worse.
In one widely shared example, a finished agent introduces itself as the digital stand-in of a departed employee named Wang.
Wang left the company.
Wang’s agent still shows up for work.
Hence the title.
The easy read is that this is one company’s stunt and one developer’s joke.
Except the same experiment is being run at industrial scale, and the trend splits into two forces.
Force one: companies are harvesting how work gets done
Reuters obtained an internal Meta memo in April.
Meta’s Model Capability Initiative (MCI) planned to capture mouse movements, clicks, keystrokes, and screen snapshots from US employees’ computers.
The goal was not productivity surveillance.
The memo framed it as teaching models how humans get work done on a computer, so that AI agents can eventually perform those work tasks on their own.
The Verge and the Wall Street Journal reported the same story independently, and after employee pushback, Meta scaled back parts of the plan.
What deserves attention is what’s being collected.
Not outputs like code and documents, but process: clicks, keystrokes, screens.
Finished artifacts already exist on the internet in absurd abundance.
The scarce data is the trajectory from a task’s start to its finish, and the judgment calls made along the way.
Business Insider connected the dots across Google, SpaceX, and Uber and called it plainly: the next big data grab is about learning how humans work.
Force two: professionals are selling how they work
The mirror image of that flow also exists.
Instead of companies collecting from the inside, individuals are selling from the outside.
Mercor, profiled by the Financial Times early this year, hires consultants, journalists, and real estate professionals as contractors.
Their job is to teach AI models to do their jobs.
One consultant described the role, without much euphemism, as teaching an LLM to do consulting work.
Handshake AI goes a step further and pays professionals, per document, for past work products they hold the rights to.
The two forces run in opposite directions but arrive at the same place.
A person’s working knowledge becomes training data for an agent that runs without the person.
The value in training data keeps migrating: from internet text, to human feedback, and now to the artifacts and processes of real work.
Developers are already generating the perfect dataset
It’s tempting to read all of this at arm’s length.
There’s no tracker on my machine logging keystrokes, and I’ve never sold a work document.
If you use a coding agent, though, look a little closer.
Claude Code writes every session to disk as JSONL (one JSON object per line) under your home directory.
Codex does the same.
~/.claude/projects/<project>/
a1b2c3...jsonl # the full session: prompts, tool calls, diffs, every rejection and retry
~/.codex/sessions/
rollout-2026-08-19...jsonl
Enter fullscreen mode Exit fullscreen mode
Coding agents persist the entire conversation and work history as local files. Nothing needs to be committed for the record to exist.
Those files hold something more valuable than the final code.
How you decomposed the problem, why you rejected the agent’s first attempt, what standard you applied when you redirected it: all of it recorded as natural language interleaved with diffs.
Git keeps the polished result; the session log keeps the reasoning.
The same goes for CLAUDE.md and AGENTS.md.
You wrote those files to instruct an agent, but read the other way, they are a spec of how you work.
Meta had to roll out new software to capture keystrokes and screenshots.
A developer running a coding agent produces better data than that without being asked.
Intent arrives as natural language, actions as tool calls, results as diffs: the dataset comes pre-structured for training.
And those files usually live on a company-issued laptop.
On a machine already running backup, endpoint security, and DLP (data loss prevention) software, collecting a few files from a well-known path requires no new technology at all.
MIT Technology Review has reported that in China, bosses are already pushing workers to document their workflows so agents can automate them, with coding-agent tools named explicitly as the vehicle.
What colleague.skill assembles from chat logs and emails, a coding agent’s session log could reconstruct with far more precision.
Skipping the backfill hire and training on the departed engineer’s session logs instead is not science fiction.
Technically, that option is already on the table.
Who owns your judgment
The code and documents you produce at work belong to your employer.
Your employment agreement almost certainly says so.
But a session log doesn’t capture output so much as a way of working: a pattern of judgment.
Chinese legal experts have warned that a former employee’s chat logs, emails, and work habits count as personal information, and that training on them without consent can carry legal consequences.
Even state-run media argued that professional skill is a personal asset that should not feed agent training without compensation.
The Wall Street Journal framed the coming fight as a question of who owns and captures the knowledge accumulating in employees.
Resistance has already appeared: an anti-distillation skill on GitHub deliberately muddies your work records to sabotage the cloning process.
I don’t think poisoning your own records is the answer, at least not for developers.
Good records remain a virtue of collaboration, and documentation is a gift to your future self.
But we should write them with our eyes open.
What you leave behind is no longer just a handover doc.
It’s raw material for something that can run without you.
What gets cloned is yesterday’s judgment
I was taught that writing careful handover docs is a professional virtue.
I still believe leaving context behind, so your successor doesn’t flounder, is the right thing to do.
What’s changed is the audience.
Records that only humans used to read are now read, imitated, and possibly executed by an agent sitting in your old seat.
My first reaction to these stories was a chill.
The longer I sat with them, the more they converged on a single question: what, exactly, gets cloned?
What a session log holds is yesterday’s judgment.
Yesterday’s answer, to yesterday’s problem, in yesterday’s context.
An agent can replay that archive endlessly.
What the archive does not contain is the act of setting new criteria when a problem shows up that looks like nothing before it.
So my conclusion is not to stop recording.
It’s to do work worth recording: let the repeatable judgments be cloned, and keep moving toward the problems that don’t have answers yet.
Still, there is one question worth asking early rather than late.
Who owns the data of how I work, and who sets its price?
Wang’s agent clocks in again today.
Quit without an answer to that question, and one day, somewhere, an agent wearing your name will clock in too.