Quantifying Multi-Agent Systems: When and Why They Work Best

January 28, 2026

Yubin Kim, Research Intern, and Xin Liu, Senior Research Scientist, Google Research

In the rapidly evolving landscape of AI, multi-agent systems are gaining traction for their ability to handle complex, multi-step tasks. However, the assumption that "more agents are better" has been a common heuristic without much empirical backing. In our latest research, we delve into the quantitative scaling principles of these systems, evaluating 180 different agent configurations to uncover when and why multi-agent setups excel or falter.

Key Findings

• Parallelizable Tasks: Multi-agent coordination significantly boosts performance on tasks that can be broken down into parallel sub-tasks.
• Sequential Tasks: Conversely, adding more agents degrades performance on sequential tasks where each step depends heavily on the previous one.
• Predictive Model: We developed a model that accurately predicts the optimal architecture for 87% of unseen tasks.

The Complexity of Multi-Agent Systems

AI agents are increasingly being deployed in real-world applications, from coding assistants to personal health coaches. Unlike traditional machine learning models, which focus on single-shot predictions, agents must manage sustained, multi-step interactions where a single error can have cascading effects. This introduces a new layer of complexity and shifts the focus from mere accuracy to overall system performance.

Heuristics vs. Empirical Evidence

The assumption that "more agents are better" has been widely accepted, with some studies like "More Agents Is All You Need" reporting improved performance with increased agent count. However, our research challenges this notion by providing empirical evidence through a controlled evaluation of 180 different configurations.

Controlled Evaluation

• Task Types: We evaluated both parallelizable and sequential tasks.
• Agent Configurations: Ranging from single-agent setups to systems with multiple specialized agents.
• Performance Metrics: We measured task completion time, error rates, and overall efficiency.

Key Observations:

• Parallel Tasks:

  • Performance Boost: Adding more agents led to a significant reduction in task completion time and error rates.
  • Optimal Configuration: The optimal number of agents varied depending on the specific task but generally increased with task complexity.

• Sequential Tasks:

  • Performance Degradation: Increasing the number of agents resulted in higher error rates and longer task completion times.
  • Coordination Overhead: The overhead of coordinating multiple agents outweighed any potential benefits, leading to suboptimal performance.

Predictive Model

To address the challenge of identifying the best architecture for a given task, we developed a predictive model. This model:

• Input: Task characteristics (e.g., parallelizability, complexity).
• Output: Optimal number and type of agents.
• Accuracy: The model correctly identified the optimal configuration for 87% of unseen tasks.

Implications for Practitioners

Our findings have significant implications for practitioners designing multi-agent systems:

• Task Analysis: Carefully analyze the nature of the task to determine if it is parallelizable or sequential.
• Agent Design: Tailor the number and type of agents based on task characteristics to avoid unnecessary overhead.
• Performance Evaluation: Continuously monitor performance metrics and adjust configurations as needed.

Conclusion

Multi-agent systems offer powerful capabilities for handling complex tasks, but their effectiveness depends heavily on the nature of the task and the system architecture. By providing quantitative scaling principles and a predictive model, our research aims to guide practitioners in designing more efficient and effective multi-agent systems.