Synthetic Computers at Scale for Long-Horizon Productivity Simulation
| Authors | Tao Ge et al. |
| Year | 2026 |
| Field | AI / Agents |
| arXiv | 2604.28181 |
| Download | |
| Categories | cs.AI, cs.CL, cs.LG |
Abstract
Realistic long-horizon productivity work is strongly conditioned on user-specific computer environments, where much of the work context is stored and organized through directory structures and content-rich artifacts. To scale synthetic data creation for such productivity scenarios, we introduce Synthetic Computers at Scale, a scalable methodology for creating such environments with realistic folder hierarchies and content-rich artifacts (e.g., documents, spreadsheets, and presentations). Conditioned on each synthetic computer, we run long-horizon simulations: one agent creates productivity objectives that are specific to the computer's user and require multiple professional deliverables and about a month of human work; another agent then acts as that user and keeps working across the computer -- for example, navigating the filesystem for grounding, coordinating with simulated collaborators, and producing professional artifacts -- until these objectives are completed. In preliminary experiments, we create 1,000 synthetic computers and run long-horizon simulations on them; each run requires over 8 hours of agent runtime and spans more than 2,000 turns on average. These simulations produce rich experiential learning signals, whose effectiveness is validated by significant improvements in agent performance on both in-domain and out-of-domain productivity evaluations. Given that personas are abundant at billion scale, this methodology can in principle scale to millions or even billions of synthetic user worlds with sufficient compute, enabling broader coverage of diverse professions, roles, contexts, environments, and productivity needs. We argue that scalable synthetic computer creation, together with at-scale simulations, is highly promising as a foundational substrate for agent self-improvement and agentic reinforcement learning in long-horizon productivity scenarios.
Engineering Breakdown
Plain English
This paper introduces a scalable system for generating realistic synthetic computer environments populated with authentic file hierarchies, documents, spreadsheets, and other productivity artifacts. The key innovation is using two cooperating agents—one to generate realistic long-horizon work objectives (requiring ~month of human effort) and another to execute those tasks across a simulated desktop environment. This enables large-scale synthetic data generation for training productivity agents without requiring humans to manually create thousands of realistic computer setups. The approach bridges a critical gap in agent training: most prior work uses simple, abstract environments, but real productivity work depends heavily on specific, user-tailored file structures and document content.
Core Technical Contribution
The technical novelty is a scalable synthetic environment generation pipeline that combines procedural folder hierarchy creation with content-rich artifact synthesis, coupled with a dual-agent simulation framework. Rather than manually curating training scenarios or using simplified task environments, the authors automated the creation of realistic computer workspaces by having an agent generate context-aware work objectives for each synthetic environment, then having another agent execute multi-step workflows to solve them. This represents the first systematic approach to scaling productivity simulation beyond toy domains—previous work either used hand-crafted environments with limited diversity or simplified task formulations that don't capture real workflow complexity. The key insight is that authentic context (folder structure, document relationships, user-specific organization) can be synthesized programmatically and then used to anchor realistic task generation.
How It Works
The system operates in two main phases: environment synthesis and simulation. First, for each synthetic computer, the pipeline generates a realistic folder hierarchy with domain-appropriate structure (e.g., project folders, archive directories, reference materials) and populates it with content-rich artifacts like documents with actual text, spreadsheets with formulas and data, and presentations with structured content. These artifacts are created to be internally consistent and contextualized—spreadsheets reference documents, presentations cite data sources, mimicking real professional environments. Second, a task-generation agent examines the synthetic computer's state and creates long-horizon productivity objectives (requiring ~month of effort) that are specifically tailored to that environment's content and structure—for example, "synthesize Q4 analysis from these quarterly reports and existing market research." Finally, an execution agent attempts to complete these objectives by interacting with the desktop environment, creating a rich trace of multi-step work. The architecture uses one LLM-based agent for environment-aware task generation and another for execution, with the synthetic computer serving as the ground truth state.
Production Impact
For teams building AI productivity agents, this work solves the critical data bottleneck: training agents for real computer interaction previously required either hand-crafted datasets (expensive and limited diversity) or oversimplified benchmarks that didn't transfer to production. By automatically generating diverse, realistic environments with context-aware tasks, teams can now create large-scale synthetic training datasets tailored to their target user workflows without manual annotation. Production impact includes: (1) bootstrapping agent training with domain-specific synthetic data before collecting human feedback, (2) testing agent behavior across thousands of realistic scenarios before deployment, and (3) generating diverse failure cases for robustness evaluation. The trade-off is significant: generating high-quality synthetic environments requires running LLMs in a loop (task generation + artifact synthesis), so creating a dataset of 10,000 environments could cost thousands of dollars and require weeks of compute. Integration requires building or adapting a desktop simulation layer that the agent can interact with and that faithfully represents the synthetic file system and artifacts.
Limitations and When Not to Use This
The paper assumes that agent behavior trained on synthetic environments will transfer to real user computers, but real environments have messier hierarchies, older conventions, and idiosyncratic organization patterns that may not be captured by programmatic generation. The synthetic tasks, while long-horizon, are still generated by an LLM with its own biases and may not reflect the full diversity of real human work—edge cases, unusual breakdowns, and context switches may be underrepresented. Artifact quality and realism are constrained by generation speed: creating deeply coherent, multi-document project contexts requires either expensive multi-turn generation or risks creating internally inconsistent synthetic data that agents may exploit. The approach also doesn't address the sim-to-real gap for agent actions—successfully completing a task in a desktop simulator doesn't guarantee success on real systems with different UI frameworks, latency, or unexpected states. Follow-up work needed includes validation that agents trained on this data actually improve on real productivity tasks, and techniques to detect and correct distributional mismatch between synthetic and real environments.
Research Context
This work builds on the emerging field of agent benchmarking and simulation (e.g., WebArena, ScreenSpot) but extends the focus from web-based tasks to offline productivity work with rich local context. It connects to recent work on synthetic data generation for instruction following and multi-step reasoning, applying those techniques to the domain of desktop automation. The research opens a new direction: moving from evaluating agents on curated benchmarks toward training them on scalable, diverse synthetic environments that better reflect real-world complexity. This aligns with broader trends in scaling agent training data (similar to how synthetic data has enabled vision and NLP model scaling), and positions productivity automation as a key testbed for long-horizon planning and grounding in complex environments.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::