Title: Multi-Agent Computer Use

Abstract:

Current implementations of Computer Use Agents (CUAs) predominantly rely on a single, serial agent architecture. This approach is ill-suited for complex, long-horizon tasks that require task decomposition, parallel processing, and dynamic re-planning in response to new information. This paper advocates for a shift toward the development and evaluation of Multi-Agent Computer Use (MACU) systems. By prioritizing coordinated planning and parallel execution, these systems address numerous limitations inherent in single-agent CUAs.

We introduce a general multi-agent framework where a central manager model structures computer use tasks as a Directed Acyclic Graph (DAG). This structure encodes specific goals and dependencies for various subagents. During each iteration, the manager assigns parallel CUA subagents to execute nodes located on the ready frontier of the DAG. As subagents report new findings, the manager continuously updates the DAG by adding, canceling, or rewriting nodes. This architecture treats the partially observable nature of computer environments as a primary challenge; crucial information that downstream agents might not be able to re-observe is preserved and transmitted through the manager and the DAG structure.

Our results show that MACU consistently outperforms strong single-agent baselines, achieving improvements ranging from $3.4\%$ to $25.5\%$ across desktop benchmarks (OSWorld) and web navigation datasets (Online-Mind2Web, WebTailBench, and Odysseys). The system also demonstrates superior test-time scaling and successfully resolves complex, long-horizon tasks that typically cause single-agent CUAs to stall. Notably, on the Odysseys benchmark, which focuses on long-horizon web navigation, MACU reduced the average wall-clock time for task completion by approximately $1.5 \times$, proving its ability to accelerate traditionally sluggish CUA workflows. These findings underscore multi-agent coordination as a vital pathway for scaling computer use agents to operate more effectively and productively over extended periods. All code and interactive visualizations are available at https://jykoh.com/multi-agent-computer-use.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC