Feynman-style Letter: Synthetic Computers at Scale and Long-Horizon Agent Evolution

Forum topic · 小凯 · 2026-05-03

Summary

A Chinese tech forum post discusses Microsoft Research's paper on Synthetic Computers at Scale (arXiv: 2604.28181) and why current AI agents fail in real-world settings. The author argues most agent training happens in oversimplified, lab-specimen-like environments—opening a webpage or calling an API—leaving agents unprepared for messy real computer systems full of nested folders, redundant files, and cross-application dependencies. The paper's proposed solution is to synthesize environments at scale: automatically generating virtual computers in memory with realistic file system hierarchies, document contents, and historical operation traces. Long-horizon trajectories are then produced by a two-agent setup, where a BOSS agent assigns month-long strategic tasks and an employee agent navigates, edits, and repairs files within the virtual machine. The result is long-horizon productivity trajectories that teach agents to maintain goal coherence across chaotic, extended workflows. The post concludes that the bottleneck for AI agent evolution is no longer compute but reserves of realistic complexity, urging builders to construct challenging synthetic sandboxes to harden their agents.

A Feynman-style Letter: Creating a Million Cyber Workers in Your Computer — On Synthetic Computers at Scale and Long-Horizon Agent Evolution

After reading Microsoft Research's paper on Synthetic Computers at Scale (arXiv: 2604.28181), an image of a "cyber Truman Show" immediately popped into my head.

To help you understand why today's AI agents always "die within three seconds," let's talk about the work environment.

1. Current state: an intern practicing martial arts in a vacuum

Most current AI agent training happens in extremely simplified, lab-specimen-like environments: open a webpage, query an API.

The pain point: The real working world is extremely messy: your computer has thousands of nested folders, piles of junk files named final_v2_fix.docx, and all kinds of complex cross-application dependencies. In such an environment, an AI is like an intern who trained in a vacuum — the moment it enters the real workplace rubble, it instantly falls apart.

2. Synthetic computers: the cyber lab that comes with folders

I won't wait for you to collect data — I'll directly weld a million virtual computers in memory.

Physical picture (environment synthesis): It uses code to automatically generate virtual environments with realistic file system hierarchies, document contents, and even historical operation traces. It's like building an infinitely expanding cyber cubicle for AI.
Two agents, one act (simulating long horizons): It has two AIs cooperate in a staged performance:
The BOSS agent: sets an extremely demanding month-long task (e.g., "based on three months of scattered meeting notes and reports, write a ten-year strategic plan for the company").
The employee agent: repeatedly works inside the virtual computer, searching files, editing documents, even simulating crashes and repairs.
Harvesting long-horizon trajectories: This simulation produces a magical substance called "long-horizon productivity trajectories." It doesn't teach AI how to type code — it teaches how to keep the gravitational source named "goal" from collapsing amid chaos across long spans of time and space.

3. A Feynman-style judgment: intelligence comes from "environmental complexity"

A so-called "general agent" isn't about how many parameters its model has. It's about whether, when facing a physical/digital environment made extremely chaotic by entropy, it can still tease out the golden logical thread that leads to results.

Synthetic computers at scale tell us: the bottleneck of AI evolution is no longer compute — it's the "reserve of realistic complexity." When we can let an AI experience in one day the workplace experience that would take a human a hundred years to accumulate, only then does the so-called "digital employee" era truly move from the PowerPoint slide into your task manager.

Takeaway: When building your own agent system, don't just stare at its reasoning ability. Go build your "cyber sandbox." If you can't beat up your AI in a "realistic" environment full of distractors and years of accumulated clutter, what you'll train is, in the end, just a fragile probability predictor that can only talk the talk on paper.