Construction and Optimization of Multi-Agent Systems: Exploring Capability Partitioning, Performance Enhancement, and Scalability Patterns

[Mitchellhk]

Multi-agent systems can decompose complex tasks into multiple specialized agents that collaborate to accomplish them. They represent a crucial form of AI implementation today and are gradually transforming human-machine interaction.

At the system construction level, three core challenges remain central:
① Capability Allocation: When building a minimum viable multi-agent system, how should capabilities be distributed among distinct agent roles?
② Pathways for Performance Scaling: To double a multi-agent system's capability, should the focus be on breakthroughs in underlying large models or on advancing the multi-agent collaboration framework itself?
③ The Scaling Law of Multi-Agent Systems: While significant scaling laws exist in large-model domains, do similar scaling patterns apply to multi-agent systems? If not yet observed, what are the primary bottlenecks?

Recently, during a technical roundtable organized by OpenAgents, experts engaged in in-depth discussions around these three critical questions.

Capability Partitioning: How to Assign “Responsibilities” to Agents?
The issue of agent capability partitioning is a critical challenge when building a minimum viable multi-agent system, and it is the primary factor determining the system's efficiency, maintainability, and scalability.

Andy Zhang, Solutions Architect at AWS Startup Team, draws from his application development experience to advocate for adhering to the Single Responsibility Principle during agent capability allocation. Each agent should focus solely on one core capability domain, enabling specialized division of labor. Clearly defining the scope of responsibilities between agents establishes distinct boundaries, thereby avoiding functional redundancy, achieving complementary capabilities, and enhancing system efficiency. Here, Andy provides three examples of hierarchical segmentation:

• Cognitive Level: Agents responsible for perception, reasoning, and execution are separated
• Temporal Dimension: Real-time response agents are separated from long-term planning agents
• Task Complexity:
  • Simple tasks use lightweight models (e.g., DMC-P)
  • Complex tasks use large models (e.g., Cloud)

Fan Wendong, a core member of Camel AI, further elaborates on the Single Responsibility Principle, arguing that agent capability allocation is fundamentally a process of “scenario clustering”—that is, grouping and assigning agents based on specific business scenarios. Using information retrieval as an example, he explains that agents should be built for corresponding business scenarios, with appropriate tools (browsers, search engines, etc.) configured beneath them. When assigning responsibilities to agents, controlling granularity is crucial. We need to perform clustering with appropriate granularity for task scenarios, rather than defining a separate agent for every distinct scenario. This prevents an excessive number of agents and collaboration bottlenecks caused by over-segmentation.

Regarding practical allocation strategies, Andy suggests considering two dimensions: task complexity and data type. For task complexity: lightweight agents can handle low-complexity tasks with rapid response; medium-complexity tasks should be managed by domain-specialized agents; high-complexity tasks require agent collaboration, such as through the OpenAgents framework. For data type: agents can manage databases and API calls for structured data. For unstructured data like text, images, speech, or even cross-modal data, agents should be tasked with integrating such multimodal information.

Regarding the overall design of multi-agent systems, Andy proposes three key considerations: First, minimizing communication overhead by standardizing message formats and asynchronous mechanisms across the system. Second, fault isolation to ensure single points of failure do not impact the entire system, enabling degradation and backup. Third, dynamic scalability requires consideration of plug-in architecture, load balancing, and mechanisms for dynamic capability redistribution.

Building on this, Andy draws an analogy to the MVC pattern in traditional web development, proposing a three-tier architecture reference for constructing a minimum viable multi-agent system: the input agent handles user intent understanding, the core business agent executes primary business logic, and the output coordination agent integrates results to generate the final response.

Performance Improvement Path: Model Breakthrough or Framework Advancement?
To double the performance of multi-agent systems, should we rely on breakthroughs in the underlying large models themselves, or on advancements in multi-agent collaboration frameworks?

Fan Wendong believes that the foundational capabilities of the models and the framework capabilities complement each other. Framework improvements are constrained by the upper limit of model capabilities, while engineering optimizations of the framework can enhance overall system efficiency without changing the model's capabilities, pushing the model closer to its capability ceiling. At this stage, frameworks serve to fully unlock model potential, while model advancements themselves reduce the burden on framework design.

Raphael Shu, founder of OpenAgents, reflects on his entrepreneurial journey, noting he too pondered whether multi-agent collaboration requires specialized models trained on dedicated datasets. Considering current collaboration practices, he argues the approach should be scenario-specific. For certain pipeline-style tasks, a clear framework defining the workflow is sufficient to achieve efficient collaboration. However, in scenarios requiring dynamic games, debates, or deep multi-agent collaboration, the collaborative capabilities of the models themselves become critical. Raphael added that the industry has yet to clarify whether “collaborative capability” should be optimized as a core model feature or as a specialized framework layer—a topic currently under academic discussion.

Raphael particularly emphasized the advantages of multi-agent local games, noting that games can significantly enhance system performance. He cited an example from an office layout optimization experiment conducted by one of his doctoral students. In the experiment, they divided a 400-square-meter office space accommodating 15 people into three optimization dimensions—safety, space utilization, and comfort. Three agents were each assigned to focus on one dimension and engaged in a game on the same layout.

Results showed that a single agent achieved only about half the layout score of the multi-agent solution after 20 reasoning steps. Moreover, the multi-agent system's total score improved by at least 160% compared to the single agent within the same number of reasoning steps, with scores continuing to rise as the number of game rounds increased. This resembles a scenario in human society where three distinct experts compete against each other; they are highly likely to achieve suboptimal solutions within their respective specialties. However, if they collaborate, mutual deference and compromise often prevent them from reaching the optimal solution. This demonstrates that the cooperative game mechanism overcomes the limitation of single-agent reasoning, which cannot guarantee finding the global optimum. It also shows that even with an unchanged underlying language model, the cooperative game mechanism itself can deliver significant performance gains.

The Question of Scaling Laws: Do Multi-Agent Systems Follow Scaling Laws?
Scaling laws have become well-established in the field of large-scale models. Do multi-agent systems also exhibit similar patterns where scale brings about capability leaps? If not yet observed, where do the primary bottlenecks lie?

Fan Wendong maintains this remains an open question, noting that the relationship between agent count and efficiency is highly context-dependent. Researchers proposed identifying scaling laws for agents as early as 2023, yet recent studies from diverse perspectives have yielded contradictory conclusions: increasing agents improves performance in some tasks while worsening it in others due to rising system complexity.

He referenced recent findings from the paper “The Science of Scaling Law Agent Systems,” which highlights topological structure as the critical determinant of performance gains: centralized multi-agent systems suppress error propagation, whereas loosely coupled structures amplify errors. For instance, under supervised conditions, independent agents can magnify errors by over tenfold, while centralized systems limit amplification to approximately three to four times.

Additionally, increasing agent numbers yields varying effects across different scenarios. For instance, linear, simple tasks may actually perform worse with multi-agent systems. Practical testing in real-world business scenarios is needed to determine optimal agent counts. Complex scenarios might require over a dozen agents, while simple ones could be solved with just two or three agents at relatively low overall cost.

Raphael Shu argues that as human utilization of AI agents increases, many who require agents may not possess engineering skills themselves. Instead, they are individuals with deep domain expertise but limited technical capabilities. Such users can leverage today's robust model support and application tools to create agents without coding, empowering their work and unlocking greater value. For instance, a domain-specific agent developed by a financial analyst could save them two hours weekly—generating at least $1,000 in value. Future waves of specialized agents across verticals (e.g., legal review) will emerge. These agents, embodying domain expertise and workflows, may proliferate in both quantity and quality, potentially spawning a novel online ecosystem governed by a unique Scaling Law.

conclusion:
It is evident that multi-agent systems are shifting from the mere aggregation of individual agent capabilities toward the construction of collaborative networks. Developers must not only focus on model selection and prompt engineering but also adopt a systems-thinking approach to strike a balance between capability partitioning, collaborative frameworks, and scaling strategies.

As Raphael Shu summarizes: “The most valuable multi-agent systems are rarely conceived by engineers alone. Instead, they emerge when domain experts—financial analysts, lawyers, designers—integrate their specialized knowledge and workflows into intelligent agents.” Moving forward, the evolution of multi-agent systems will be driven by three converging forces: breakthroughs in underlying models, innovations in collaborative frameworks, and vertical domain integration. This convergence will ultimately forge a decentralized, complementary, and dynamically assembling network of agent services.