Multi-Agent Computer Use

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.

• The paper’s main contribution is a manager-driven, DAG-based multi-agent framework that improves success rates by up to 25.5% on complex tasks.
• It employs continuous replanning and parallel execution across isolated VMs to efficiently coordinate and recover from task failures.
• Experimental results demonstrate significant performance gains over single-agent setups, especially on benchmarks like Odysseys with long-horizon tasks.

Multi-Agent Computer Use: A Technical Analysis

Motivation for Multi-Agent Computer Use

The landscape of Computer Use Agents (CUAs)—autonomous agents capable of interacting with GUIs in complex environments—has largely centered on single-serial agent architectures. This single-agent paradigm reveals significant deficiencies on long-horizon, compositional, and partially observable tasks, which are tractable only under decompositional planning, parallel execution, and continual replanning. The "Multi-Agent Computer Use" (MACU) paradigm addresses these deficiencies by recasting computer use as a partially observable, dynamically changing planning problem best managed by multi-agent systems operating over a structured task decomposition.

Figure 1: The MACU framework executes a task by organizing it as a DAG, launching parallel subagents at the ready frontier, and continuously replanning based on newly collected information.

The central innovation of MACU is the deployment of a manager agent that maintains a dynamic Directed Acyclic Graph (DAG) of subtasks. This manager incrementally decomposes, assigns, revises, and aggregates subproblems marshaled to parallel CUA subagents. Each subagent operates in an isolated virtual machine (VM), running independently and reporting back to the manager, enabling robust handling of stateful, partially observable environments—where downstream subagents may only receive state via explicit information passing.

MACU Framework and Execution Flow

The MACU system consists of a two-level agent architecture:

1. Manager Agent: An LLM-based planner that decomposes tasks into a DAG of subtasks, dispatches parallel CUA subagents to ready nodes, and perpetually replans the DAG as findings return. The manager is responsible for initial decomposition, adaptive replanning, explicit file management across VMs, subtask continuation, and final aggregation.
2. CUA Subagents: Homogeneous actors (often using ReAct-style loops) that execute instructions in VM sandboxes, observe screengrabs, perform reasoning and action cycles, and produce structured outputs as instructed.

At initialization, the manager synthesizes a decomposition into executable subtasks, identifying dependencies for correct sequencing. As subagents report results—files, browser states, discovered obstacles, or partial completions—the manager selectively replans, adding retries, structural variants, or altering downstream instructions. Notably, the manager can also attach archived files from one subagent's output to another's input to enable complex multi-step workflows (e.g., producing a document in one subtask, editing or using it in another).

Task completion is recognized by the final aggregation node in the DAG, which combines results and, where necessary, designates a final VM state as the deliverable.

Figure 2: Example of MACU managing a travel planning task, parallelizing hotel and flight planning, performing retries on blocked branches, and aggregating results for final output.

Experimental Results: Empirical Performance and Scaling

The MACU framework is evaluated on four major benchmarks: OSWorld (desktop execution), Online-Mind2Web (live web navigation), WebTailBench-v2 (multi-step and compositional web tasks), and Odysseys (long-horizon realistic web agent tasks).

Strong single-agent baselines using Qwen3.6-27B as the CUA are compared against MACU with two configurations: manager as either Qwen3.6-27B or Claude Opus 4.6. The manager's capacity for dynamic replanning and DAG manipulation is shown to be indispensable for success on complex tasks.

On all benchmarks, MACU demonstrates absolute improvements in success rate ranging from 3.4% (Online-Mind2Web) up to 25.5% (Odysseys) relative to the single-agent setup. Notably, Odysseys—characterized by extended planning horizons and high compositional complexity—sees success rates quadruple under MACU. For OSWorld and Odysseys, the multi-agent paradigm yields substantial wall-clock time reductions (1.47× speedup on Odysseys), attributable to parallelism.

Figure 3: MACU's success rate increases with the number of CUA actions executed, outpacing single-agent setups as more compute budget is applied.

MACU also demonstrates superior utilization of inference-time compute: scaling the available CUA actions or parallelism yields monotonic success rate improvements, with single-agent baselines plateauing prematurely.

Ablations: Importance of Continuous Replanning and Parallelism

Ablation studies probe the contribution of the dynamic replanning budget ($B$), parallelism parameter ($N$), and both worker and manager model strengths.

• Replanning Budget: Allowing only the initial static plan ($B=0$) results in poor success rates. Increasing $B$ (i.e., enabling the manager to revise the DAG mid-execution) drastically elevates success rates—underscoring the necessity of continuous adaptive planning in partially observable settings.
• Parallelism ($N$): On decomposable tasks (e.g., Odysseys), increasing $N$ yields near-linear wall-clock time reductions and further improves average rubric-based scores due to increased task coverage.
• Manager/Worker Capacity: Gains compound as both manager and CUA model strengths improve, but the benefits of a strong manager (Opus 4.6) are especially pronounced even for weaker subagents, indicating that the coordination policy is a substantial bottleneck in serial approaches.

  Figure 4: Impact of the maximum number of parallel subagents ($N$) on wall-clock time, success rate, and rubric score in Odysseys tasks.

Comparative analysis against naive pass@$k$ baselines—repeated restarts of single-agent attempts without graph structure or information aggregation—shows MACU leverages compute budget more efficiently, even without access to ground-truth evaluators.

Figure 5: MACU accelerates success rates over the single-agent baseline across all difficulty levels in Odysseys and Online-Mind2Web.

Structural Patterns and Domain Impact

Task decomposition in MACU is flexible, supporting a spectrum of structural patterns:

• Serial chains for inherently sequential, single-step tasks.
• Map-Reduce for compositional or information-gathering tasks, enabling parallel exploration and aggregation.
• Retry chains and structural variants for error recovery, blocking state, or exploration of alternate strategies.
• State-passing via explicit file management for multi-stage application workflows.

These patterns are most leveraged in benchmarks with compositional or long-tailed objectives (e.g., Odysseys, WebTailBench-v2), where the partial observability and non-monotonic progress of single-agent CUAs limit achievable success rates. OSWorld and Online-Mind2Web, with a higher prevalence of atomic GUI steps, admit smaller absolute gains.

Empirically, MACU’s largest delta in success rate emerges in decomposable, parallelizable, and long-horizon tasks—consistent with theoretical predictions for DAG-based parallel execution in partially observable environments.

Theoretical and Practical Implications

MACU provides an extensible, manager-centric interface for orchestrating independent CUA subagents. Unlike specialized multi-agent and co-agent systems, MACU is agnostic to the specific CUA model—enabling straightforward integration of future advances in agent capabilities. Manager-subagent decoupling also facilitates modular research on planning, scheduling, and error recovery policies, without retraining or altering low-level execution agents.

From a theoretical standpoint, MACU demonstrates that multi-agent coordination over an explicit task DAG is not merely system engineering—it unlocks demonstrable scaling advantages in environments characterized by partial observability, state mutability, and frequent dead-ends. Notably, MACU’s ability to manage file artifacts, resume state via VM inheritance, and exploit structural variants in execution policy provides a new axis of agentic competence not captured by monolithic, serial, or “chain-of-thought” type decision protocols.

Looking ahead, several axes of improvement emerge:

• Learning-Based Manager Training: Fine-tuning or RL for managers trained jointly with subagents could improve decomposition and replanning quality.
• Adaptive Budgeting and Resource Management: Dynamic selection of $B$ and $N$ based on real-time estimations of decomposability and bottleneck detection.
• Richer Cross-Agent Communication: Sharing abstracted state or partial knowledge between subagents beyond strict file passing, potentially leveraging multi-agent communication primitives.
• Scalability and System Monitoring: Advanced orchestration and monitoring tools to manage large pools of agents and VMs for practical deployments.

Conclusion

MACU defines a principled, empirically validated framework for multi-agent computer use that elevates CUA capabilities on complex, stateful, and partially observable tasks, outperforming strong single-agent baselines. The system explicates the role of a manager for dynamic task decomposition and robust error recovery—mechanisms critical as real-world deployments extend to longer horizons and greater task complexity. Multi-agent coordination through structured DAGs is poised to become a foundational paradigm in the next generation of autonomous software agents.