Definition
Computer-use agents drive a real computer interface — mouse, keyboard, screenshots — rather than calling structured APIs. They can in principle do anything a human can on a screen, which is exactly why they’re also the most worrying class of agent to deploy without good oversight.
Episodes covering this
Worth reading next
Papers we haven't done a deep dive on yet, but would recommend on this topic.
- OSWorld: Benchmarking Multimodal Agents for Open-Ended Tasks in Real Computer Environments
- WebArena: A Realistic Web Environment for Building Autonomous Agents
- GUI Odyssey: A Comprehensive Dataset for Cross-App GUI Navigation on Mobile Devices
- AndroidWorld: A Dynamic Benchmarking Environment for Autonomous Agents
- AgentTrek: Agent Trajectory Synthesis via Guiding Replay with Web Tutorials
- AppAgent: Multimodal Agents as Smartphone Users
- AgentDojo: A Dynamic Environment to Evaluate Prompt Injection Attacks and Defenses for LLM Agents
- Android in the Wild: A Large-Scale Dataset for Android Device Control