AI Papers Reader

Personalized digests of latest AI research

View on GitHub

Beyond Chatbots: Microsoft Researchers Build "Synthetic Computers" to Train the Next Generation of AI Workers

Artificial intelligence is outgrowing the chat box. While tools like ChatGPT have mastered short-form conversation, the next frontier is the “long-horizon” agent—AI that can function like a digital coworker, managing complex projects that span weeks and require navigating thousands of files.

The hurdle? Training these agents requires data that currently doesn’t exist at scale. Real-world professional work is private, messy, and deeply embedded in specific computer environments. To bridge this gap, a team of Microsoft researchers has unveiled a new methodology called “Synthetic Computers at Scale.” Instead of just giving an AI a task, they are giving the AI an entire simulated life, complete with a decade of career history, hundreds of realistic files, and demanding virtual bosses.

Building a Digital History

The researchers’ approach starts with a “persona.” Take, for example, “Margaret Elaine Forsythe,” a senior financial advisor in Denver with 16 years of experience. To train an AI to work like Margaret, the system doesn’t just create a prompt; it “hallucinates” her entire digital world.

The system generates a detailed user profile, including Margaret’s technical skill level, her preferred tools (like Excel and PowerPoint), and even her organizational quirks—perhaps she’s tidy with folder structures but forgets to archive old versions of spreadsheets. It then builds a “Synthetic Computer” for her, populating a virtual hard drive with a realistic folder hierarchy and over 100 “content-rich artifacts.” These aren’t empty placeholders; they are functional documents, spreadsheets, and PDFs that reference one another, creating a complex web of dependencies.

The Month-Long Simulation

Once the environment is set, the simulation begins. A “setup agent” assigns Margaret a month’s worth of work. For a financial advisor, this might involve refreshing a model portfolio based on new market data and onboarding a high-net-worth client.

A separate “work agent” then steps into Margaret’s shoes. Over the course of a simulated month—which equates to roughly 8 hours of real-world processing time and 2,000 individual actions—the AI must navigate the file system, analyze data, and coordinate with simulated collaborators.

These collaborators, such as a skeptical Managing Director named David Hartley, add a layer of social complexity. They might send terse emails, provide conflicting feedback, or hold back necessary information until asked. This forces the AI to demonstrate more than just technical skill; it must show “agentic” persistence, planning, and the ability to recover from failures.

Scaling Experience

In their preliminary experiments, the researchers created 1,000 of these synthetic computers. The result is a massive library of “experiential learning signals.” By reviewing these trajectories, researchers can identify exactly where an agent went wrong—such as failing to update a source spreadsheet before exporting a final report—and turn those mistakes into “skills” that improve future performance.

The results are promising. Agents trained using these synthetic experiences showed significant improvements on both specialized financial tasks and general productivity benchmarks.

The team argues that this methodology could eventually scale to millions or billions of synthetic worlds. By creating a safe, scalable “practice ground,” researchers may have found the key to moving AI from a simple assistant to a truly capable digital colleague, capable of handling the long, winding road of professional productivity.