AI research wiki
agentic-systems · 2026-05-01 · Tier 2

Synthetic Computers at Scale: Long-Horizon Productivity Simulation

Synthetic Computers at Scale: Long-Horizon Productivity Simulation

arXiv: 2604.28181 · paper · HF Tier: 2 — agent training infrastructure, synthetic data Raw: ../../raw/huggingface/2026-05-01-synthetic-computers-at-scale-long-horizon-productivity-simulation.md

TL;DR

To train productivity agents that can do real long-horizon work, you need realistic user computers — folder hierarchies, content-rich documents, simulated collaborators. This paper builds a methodology to synthesize 1,000 fully-populated computers at scale, then runs 8+ hour, 2,000+ turn agent simulations on each that produce a month's worth of fake-but-plausible work. The simulations yield "experiential learning signals" that improve agent performance on both in-domain and out-of-domain productivity benchmarks. The pitch: persona populations are abundant at billion scale, so this can in principle generate millions or billions of synthetic user worlds.

Mechanism

Two-agent loop on each synthetic computer:

┌─────────────────────────────────────────────────────────────┐
│  Synthetic computer: filesystem, docs, spreadsheets, emails │
│  populated to look like a real user's workspace             │
└─────────────────────────────────────────────────────────────┘
        │                                          │
        ▼                                          ▼
┌─────────────────────┐              ┌──────────────────────────┐
│  Objective agent    │              │  Worker agent (the user)  │
│  Generates user-    │ ───────────▶ │  Navigates filesystem,   │
│  specific objectives│              │  collaborates, produces  │
│  requiring multiple │              │  artifacts until done    │
│  deliverables and   │              │  (>2000 turns, >8 hours) │
│  ~1 month of work   │              │                          │
└─────────────────────┘              └──────────────────────────┘

The interesting design choice is decoupling objective generation from execution. The objective agent works "outside" the computer with full visibility (so objectives are realistic for that user); the worker agent works "inside" the computer (so trajectories are realistic for an agent). This separation prevents the objective agent from generating tasks that an agent would never naturally encounter.

Why this is Tier 2

The single biggest gap in agent training is realistic long-horizon trajectories at scale. ClawGym (04-30) covered medium-horizon tasks (single workflow). Persistent Agent Infrastructure (04-23) covered the runtime side. This paper covers the data side — the long-horizon, multi-day-equivalent training corpus that's been impossible to gather from real users without privacy and consent issues.

The validation is the part to keep: improvements appear on both in-domain and out-of-domain productivity evaluations. If the gains transferred only in-domain, you'd suspect overfitting to synthetic distribution. Cross-distribution improvement suggests the long-horizon training signal is real.

Connection to prior wiki

  • ClawGym (04-30) dual-route synthesis (persona-driven + skill-grounded) is the medium-horizon analog. Synthetic Computers is the long-horizon extension — same persona-driven principle, scaled to month-long workflows.
  • Persistent Agent Infrastructure (04-23) raised the question: how do you do RL on agents that mutate persistent state? Synthetic Computers is the data side of the answer — the training corpus must contain trajectories where state mutation matters.
  • OccuBench (04-16) / Claw-Eval-Live (05-01) show frontier agents fail real workflows. Synthetic Computers is the training-side bet for closing that gap.
  • LWM-based simulation is shared with OccuBench — both treat the simulated environment as a fully-rendered model of a real workspace, not a stripped-down sandbox.

Open problems

  1. Distribution shift between synthetic personas and real users. "Personas at billion scale" sounds clean, but the synthetic worker agent's behavior is still LLM-driven. If the LLM has biases (e.g., overly thorough documentation, formulaic email tone), the corpus inherits them. Cross-distribution gains don't fully rule this out.
  2. Compute cost. 8+ hours per simulation × 1,000 computers = ~10,000 GPU-hours of agent runtime just to generate the corpus. The "billions of synthetic worlds" pitch assumes inference cost continues to drop on the SemiAnalysis trajectory (also today). At current cost it is hyperscaler-scale only.
  3. Composition with verifiable grading. Synthetic Computers produces trajectories; Claw-Eval-Live grades them. The natural next paper: train on Synthetic Computers data, evaluate on Claw-Eval-Live, see whether the eval gap closes.

Research angle

The "billions of synthetic user worlds" pitch is, at base, an argument that agent training should scale via simulated environments rather than real-user logs. If true, this changes which data assets matter: the bottleneck is no longer scraping real workflows but having enough compute to simulate plausible ones. That argument is consistent with the SemiAnalysis (today) thesis on token economics — once tokens are cheap enough, simulated training corpora become preferable to real ones.