← All speakers

Bio, Work & Ideas

Robert Wendt

Conference affiliation: Cua · 2026

Robert Wendt leads infrastructure at Cua, where he builds operating-system interfaces and computing infrastructure for agents that use desktop applications. He joined the company in October 2025 and developed the Linux backend for Cua Driver, extending background desktop automation to an operating system fragmented across display servers, accessibility services, application toolkits, and window managers.

His Linux computer-use implementation combines AT-SPI accessibility data over D-Bus, accessibility actions, and X11 or XTEST input when applications require alternative interaction paths. A synthetic cursor allows an agent to work independently while a person retains control of the physical pointer. The driver supports X11 and XWayland; native Wayland support remains incomplete.

  • Reliable background desktop automation: Wendt adapts agent interactions to applications built with GTK, Qt, Electron, Chromium, and other toolkits. When accessibility services, application visibility, or desktop-session support fail, the system reports the problem explicitly instead of incorrectly claiming an action succeeded.
  • Demand-based sandbox autoscaling: Through Cua Fleets, he maintains a warm pool of desktop sandboxes that expands or contracts as GPU workers request environments. Prestarted machines absorb image-download, boot, and reset delays, allowing infrastructure to adapt during extended reinforcement-learning runs without requiring operators to choose a fixed pool size upfront.
  • Reducing idle GPU costs: Wendt treats desktop provisioning as an accelerator-utilization problem: GPUs cannot generate useful training data while waiting for large virtual-machine environments to become available. His computer-use infrastructure work shifts those delays onto comparatively inexpensive sandbox infrastructure, improving readiness while preserving clean environment resets.

Read the topics behind these talks

1 conference talk

Key ideas

Scroll to read ↓

Running agents beside a human requires more than another cursor: window-level control, trustworthy task evaluation, and infrastructure that keeps GPUs from waiting on desktops.

  • When the agent needs your desktop
    1:22 ↗
  • Cua Driver runs beside the user
    2:34 ↗
  • Give each GUI task an initial state and a success check
    6:34 ↗
  • Editing a circuit is easier than starting with a blank schematic
    8:13 ↗
  • The computer tool changes the result
    9:04 ↗
  • Try to break the evaluator before trusting its score
    9:34 ↗
  • A desktop reset can leave an expensive GPU idle
    10:50 ↗
  • Prepare sandboxes before workers need them
    13:00 ↗
  • On mobile, background work may mean tool use
    15:35 ↗

References