Your AI agents are about to practice on digital work clones

The AI industry has a new data obsession. It is no longer just text scraped from the internet or contractors rating chatbot answers. The next frontier is building digital workplaces where AI can practice doing entire jobs.

These simulated training environments, called reinforcement learning environments or RL environments, are designed to let AI agents learn through trial and error. Instead of predicting the next word in a sentence, an agent attempts a full workflow -- coding a feature, completing a business process, navigating enterprise software -- and improves based on whether it finished the task.

Two recent developments show how quickly this approach is gaining traction.

Google is in talks to invest over $1.5 billion in Mechanize, a startup that builds virtual work environments for AI training, Business Insider reported last week. The deal would bring Mechanize's talent under Google while the tech giant licenses some of its technology for model evaluation and development.

Meta, meanwhile, has been collecting employee keystrokes, mouse movements, clicks, and screen activity. The stated goal is to teach AI how people actually use computers, from keyboard shortcuts to navigating workplace software.

The first generation of large language models relied on two main data sources: vast text corpora scraped from books and the web, then human feedback from contractors who judged whether chatbot answers were good or bad. That approach produced impressive conversational AI. But it has struggled to create agents that can independently complete complex, multi-step work over hours or days.

Scale AI, one of the largest training data suppliers, says its clients are moving beyond static datasets and human preference feedback. Nearly half of the company's new AI training projects now involve RL environments that model realistic coding, computer use, and enterprise workflows, Chetan Rane, head of product for agents and RL environments at Scale AI, wrote in a blog post.

Mechanize is making an even bolder bet. The startup was launched last year by AI researcher Tamay Besiroglu with the stated goal of "full automation of the economy."

"We will achieve this by creating simulated environments and evaluations that capture the full scope of what people do at their jobs," Besiroglu and his cofounders wrote in their launch announcement. "The market potential here is absurdly large: workers in the US are paid around $18 trillion per year in aggregate. For the entire world, the number is over three times greater, around $60 trillion per year."

Mechanize's first target is software engineering. The company argues that future coding agents will learn from professional programmer examples before improving through reinforcement learning inside environments that capture the complexity of real engineering projects. As those environments improve, the same approach can expand into all kinds of white-collar work.

Reward signals are central to RL environments. "You want to be able to tell the model you did the task correctly versus incorrectly," Besiroglu told Business Insider last year. "Then you want to leverage that to reinforce the kind of patterns of behavior that resulted in it correctly performing the task."

That logic helps explain why large tech companies are suddenly interested in how their own employees work. Meta's internal announcement said its tracking software would help AI understand everyday computer tasks because agents need real examples. Uber has begun embedding top AI engineers inside departments including finance, legal, HR, marketing, procurement, and customer support to observe workflows before redesigning them around AI, calling the initiative "Agentic Pods."

The result is that the industry's newest race may not be about smarter chatbots. It is about building detailed digital versions of real workplaces, where AI agents can practice the thousands of decisions and actions that make up modern jobs.