Synthetic Computers at Scale for Long-Horizon Productivity Simulation

1. Problem Statement

As AI agents evolve from conversational assistants to long-horizon agents grounded in entire user computer environments (e.g., for coding, document creation, and project management), they require rich, user-specific context to perform realistic productivity work. However, collecting real-world trajectories at scale is prohibitively difficult due to privacy concerns regarding personal files, enterprise data, and interaction histories.

Existing synthetic data methods often fail because they generate generic, "toy" workflows that lack the complexity of real user environments. The core challenge is synthesizing not just the task, but the entire context: the specific directory structures, content-rich artifacts (documents, spreadsheets, presentations), accumulated work history, and collaborative dynamics that define a real user's computer. Without this grounding, agents cannot learn to navigate, plan, or coordinate effectively over long time horizons.

2. Methodology: Synthetic Computers at Scale

The authors propose a scalable methodology to create diverse, artifact-rich, user-specific synthetic computer environments and run long-horizon simulations on them. The process consists of three main stages:

A. Synthetic Computer Creation

This stage builds a realistic computer environment from a single "persona" through a progressive elaboration pipeline:

1. Persona-Driven User Profiles: A high-level persona (e.g., "Senior Financial Advisor") is expanded into a detailed profile including identity, occupation, career history, current projects, collaborators, and specific computer usage habits (e.g., naming conventions, file organization styles, attachment behaviors).
2. Filesystem Planning: Based on the profile, the system generates a Filesystem Policy (drive layout, naming styles, storage patterns) and a File Inventory. Crucially, it constructs a directed dependency graph where files are not independent but relate to one another (e.g., a final PDF is derived from an Excel model, which is derived from a downloaded PDF). This ensures logical consistency and realistic file evolution.
3. Artifact Creation: The system materializes the directory structure and populates it with content-rich artifacts.
   • Public files (e.g., market reports) are retrieved from the web.
   • User-specific files are generated by LLM agents using specialized tools (e.g., minimax-docx, minimax-xlsx) conditioned on the file description and its predecessors in the dependency graph.

B. Long-Horizon Productivity Simulation

Once a synthetic computer is ready, a two-agent simulation runs for approximately one month of simulated work:

1. Setup Agent: Analyzes the user profile and computer state to generate Productivity Objectives. These are complex, multi-deliverable goals (e.g., "Refresh model portfolios," "Onboard a new HNW client") requiring sustained effort. It also sets up Simulated Collaborators (managers, peers, clients) with distinct personalities, communication styles, and private reference materials.
2. Work Agent: Acts as the user, operating within the synthetic computer to complete the objectives.
   • Planning: The agent creates weekly and daily plans, breaking down tasks into specific activities (e.g., "Email client," "Revise Excel model").
   • Execution: The agent navigates the filesystem, reads existing artifacts, creates/revises files, and communicates with simulated collaborators.
   • Feedback Loop: The agent receives replies from collaborators (e.g., feedback on a draft, new data requests) and iterates on its work.
   • Duration: Each simulation averages 2,272 turns and 8.59 hours of agent runtime.

C. Experiential Learning & Skill Extraction

The simulations generate rich data from both process (trajectories, planning logs, collaboration logs) and outcome (final deliverables).

• Retrospective Analysis: A judge agent analyzes the full trajectory to identify successful patterns, failure modes, and lessons learned.
• Skill Creation: These insights are aggregated by occupation and converted into occupation-specific skills (e.g., "Financial Analyst Skill" containing rules about data integrity, version control, and regulatory compliance).
• Self-Improvement Loop: These skills are used to augment the Work Agent for future simulations, creating a cycle of continuous improvement.

3. Key Contributions

• Synthetic Computers Framework: A novel methodology for generating scalable, user-specific computer environments with realistic directory hierarchies, content-rich artifacts, and dependency graphs, moving beyond simple task generation.
• Long-Horizon Simulation Protocol: A robust simulation setup involving two agents (Setup and Work) and simulated collaborators, capable of running complex, month-long productivity workflows with over 2,000 interaction turns.
• Experiential Learning Mechanism: A demonstrated pipeline for extracting "experience" from simulation trajectories and converting it into actionable skills that improve agent performance.
• Open Resources: Release of 100 synthetic computers (50 Windows, 50 macOS) and retrospective analysis reports for 500 simulations to support further research.

4. Experimental Results

The authors instantiated 1,000 synthetic computers across diverse professions (e.g., finance, engineering, healthcare) and ran simulations on them.

• Simulation Scale:

  • Average files per computer increased from ~112 to ~197 after simulation.
  • Simulations involved ~5.5 collaborators and ~31 communications per run.
  • Artifacts were content-rich (DOCX, XLSX, PPTX, PDF), with final deliverables being substantial in size (e.g., PPTX median ~576KB).

• In-Domain Evaluation (Same Distribution):

  • Setup: 900 computers were used to extract skills; 100 held-out computers were used for testing.
  • Result: Agents equipped with occupation-specific skills extracted from the training set achieved a mean rubric score of 68.6%, compared to 61.6% for the baseline.
  • Win Rate: The skill-augmented agent outperformed the baseline in 83% of the 100 test cases.
  • Scaling: Performance improved as the number of training computers increased (from 10 to 900), indicating that broader coverage leads to more robust skills.

• Out-of-Domain Evaluation (GDPVal Benchmark):

  • The skills were tested on GDPVal, a public benchmark of 220 realistic productivity tasks (which differ significantly in structure and context from the synthetic computers).
  • Result: The skill-augmented agent (using Sonnet-based skills) significantly outperformed the baseline on the GDPVal gold set, winning 105 tasks vs. losing 67 (p=0.002). This demonstrates that the learned skills capture generalizable productivity patterns, not just simulation-specific artifacts.

5. Significance and Future Directions

• Foundational Substrate for AGI: The paper argues that scalable synthetic computer creation is a critical foundation for agent self-improvement and agentic reinforcement learning. It allows agents to practice long-horizon planning, grounding, and collaboration in safe, privacy-preserving environments.
• Scalability: Given the abundance of personas (billions possible), this methodology can scale to millions or billions of synthetic user worlds, covering diverse professions and contexts.
• Future Work: The authors suggest enhancing the realism of synthetic computers by:
  • Adding more personalized visual styles and formatting.
  • Introducing "natural noise" (e.g., abandoned files, duplicates, temporary downloads) to mimic real human behavior.
  • Creating more dynamic, multi-agent organizational simulations where collaborators have their own evolving states and workloads.

In conclusion, Synthetic Computers at Scale provides a viable path to generating the massive amounts of high-quality, grounded experiential data required to train AI agents capable of performing complex, long-horizon professional work.