In a nutshell: Multi-agent coordination with task decomposition and parallelization substantially improves computer-use agents and solves complex long-horizon tasks where single agents fail.
Researchers propose running computer-use agents not as single agents but as multi-agent systems with a manager model. This achieves 3.4–25.5 % improvements on desktop and web navigation tasks.
Currently, computer-use agents (CUAs) are deployed predominantly as single serial agents. This setup is suboptimal for complex long-horizon tasks that would benefit from task decomposition, parallel execution, and continuous replanning. A research team argues for evaluating and developing multi-agent computer-use (MACU) systems instead.
MACU systems are built on a manager model that decomposes computer-use tasks as directed acyclic graphs (DAGs), encoding dependencies and subagent objectives. At each iteration, the manager distributes parallel CUA subagents to execute nodes at the ready frontier of the DAG and continuously revises the graph (adding, removing, or rewriting nodes) as new insights from subagents arrive. The design treats the partially observable environment of computer use as a fundamental challenge: information that downstream agents may not be able to re-observe is preserved and propagated through the manager and DAG structure design.
In evaluations, MACU shows consistent improvements over strong single-agent baselines: 3.4–25.5 % on desktop benchmarks (OSWorld) and web navigation benchmarks (Online-Mind2Web, WebTailBench, Odysseys). The system exhibits better test-time scaling behavior and solves complex long-horizon tasks where single-agent CUAs get stuck. On Odysseys, a long-horizon web navigation benchmark, MACU reduces average task completion wall-clock time by approximately 1.5× and demonstrates efficiency in accelerating traditionally slow CUA pipelines.
The findings suggest that multi-agent coordination is a promising approach for scaling computer-use agents to make them more productive and enable longer-running work. Code and interactive visualizations are publicly available.
Source: arxiv.org · Published May 31, 2026
Lumi AI News — AI-assisted curation in accordance with Article 50 EU AI Act. Paraphrase and classification by Lumi News Pipeline v1.2.9.