Multi-Agent Computer Use

Researchers proposed a multi-agent computer-use framework that turns tasks into a DAG, dispatches parallel subagents, and updates plans as new information arrives.

The system outperformed strong single-agent baselines on desktop and web-navigation benchmarks, with gains of 3.4% to 25.5%.

It also completed long-horizon navigation tasks faster, showing that coordination and parallelism can improve agent performance on complex workflows.

The result suggests computer-use AI may be more practical when multiple agents plan and execute together instead of relying on one agent alone.