Multi-Agent Computer Use
2026-06-01 • Multiagent Systems
Multiagent SystemsComputation and LanguageMachine Learning
AI summaryⓘ
The authors show that using multiple computer agents working together is better than just one agent doing tasks one after another. They design a system where a manager breaks big tasks into smaller parts that different agents can do at the same time and updates the plan based on what they find. This approach helps complete complicated and long tasks faster and more reliably across different computer and web benchmarks. Their system keeps track of important information so agents don’t have to redo work, making everything smoother and quicker.
multi-agent systemstask decompositionparallel executiondirected acyclic graph (DAG)computer use agents (CUAs)long-horizon tasksplanningweb navigationpartial observabilitytask coordination
Authors
Jing Yu Koh, Ruslan Salakhutdinov, Daniel Fried
Abstract
Computer use agents (CUAs) today are primarily deployed as single serial agents. This setup is suboptimal for complex long-horizon tasks that benefit from task decomposition, parallel execution, and consistent re-planning based on new information. In this paper, we argue that we should instead move towards evaluating and building multi-agent computer use (MACU) systems. These systems, which emphasize planning and parallel execution, alleviate many of the shortcomings of single-agent CUAs. We propose a general multi-agent setup in which a manager model decomposes computer use tasks as a directed acyclic graph (DAG), encoding relevant dependencies and goals for subagents. At each iteration, the manager dispatches parallel CUA subagents to carry out nodes on the ready frontier of the DAG, and continuously revises the DAG (adding, canceling, or rewriting nodes) as new findings arrive from subagents. This design treats the partially observable environment of computer use as a first class challenge: information that downstream agents may not be able to re-observe are retained and passed forward through the manager and DAG structure. We demonstrate that MACU consistently improves over strong single-agent baselines by $3.4-25.5\%$ on desktop (OSWorld) and web navigation (Online-Mind2Web, WebTailBench, Odysseys) benchmarks, exhibits more favorable test-time scaling, and solves complex long-horizon tasks where single-agent CUAs get stuck. On Odysseys, a long-horizon web navigation benchmark, MACU improves average task completion wall-clock time by ${\sim} 1.5 \times$, demonstrating its efficacy in speeding up traditionally slow CUA pipelines. Our findings highlight that multi-agent coordination is a promising axis for scaling computer use agents to work productively for longer and more effectively. We release all code and interactive visualizations at https://jykoh.com/multi-agent-computer-use.