Chinese Tech Workers Are Being Told to Train Their Own AI Replacements

MIT Technology Review and Rest of World report that Chinese tech companies are instructing employees to build AI agents that replicate their own job functions. A GitHub project called "Colleague Skill" from Shanghai AI Lab imports chat history and files from Lark and DingTalk, then generates manuals describing a coworker's duties in a format an AI agent can execute.

Workers describe the experience as reductive. One told MIT Tech Review their work had been "flattened into modules," which is an incredibly precise description of what happens when you try to encode institutional knowledge into agent workflows. Anyone who's written documentation for an automated process knows this feeling. The messy, contextual judgment calls that make human work valuable get stripped down to decision trees and API calls.

Here's what caught my attention though. It's not working. The companies haven't successfully replaced anyone yet because the agents remain unreliable and require constant human supervision. The "Colleague Skill" approach produces agents that can follow documented procedures but break the moment they encounter anything the documentation didn't anticipate. Which, if you've worked in any organization for more than a week, you know is constantly.

I find this both reassuring and concerning for different reasons. Reassuring because it validates what I've observed in my own agent work: the gap between "follows instructions" and "exercises judgment" is enormous, and no amount of chat history import bridges it. Concerning because the intent is clear. These companies aren't experimenting. They're mandating self-replacement training. The technical limitations are a delay, not a barrier.

For builders in the West looking at this and feeling insulated: you shouldn't. The same "flatten work into modules" approach is exactly what every agentic coding tool does. When you teach Claude Code your codebase conventions, your testing patterns, your deployment workflow, you're building your own Colleague Skill. The difference is you're doing it voluntarily, and you control the output.

The uncomfortable question this story raises: what happens when the agents get reliable enough? Not this year. Maybe not next year. But the organizational intent has been expressed, the tooling is being built, and the data collection is underway.

What to do now: Look at your own work through the "Colleague Skill" lens. Which parts of your job are modular and documentable? Those are the parts agents will handle first. Invest your growth in the parts that can't be flattened: system design, taste, stakeholder judgment, the ability to know what to build, not just how to build it.