The cooperation paradox of multi-agent systems

Today, from tech giants to startups, all are promoting a new AI model: enabling multiple AI agents to work collaboratively like a human team, claiming to break through the capability ceiling of a single large model.

A research report from IDC indicates that by 2027, 60% of large enterprises will adopt collaborative agent systems, boosting business process efficiency by over 50%.

It sounds like multi - agent collaboration has paved a smooth path to stronger artificial intelligence. However, there have been some skeptical voices in the early stage of its launch. Some C - end users have reported that when dealing with complex problems, multi - agent collaboration takes longer to generate answers and consumes a significant number of tokens, and the results are not as impressive as expected.

Theoretically, when agents come together, they should achieve the effect of “1 + 1 > 2”. But why do we still encounter less - than - ideal situations in actual use?

One Brain VS One Team

Currently, there are two mainstream working modes for agents: single - agent and multi - agent collaboration. As the name suggests, a single agent means that all tasks are completed by one AI brain. Leading large models such as ChatGPT and Claude fall into the category of single agents. It is like an all - purpose Swiss Army knife, having to handle everything from answering questions to generating code on its own. This mode has a simple structure, low cost, and is easy to manage, but it has an upper limit on capabilities and a risk of single - point failure. Once the task is too complex or there is a problem with itself, the entire system may collapse.

To solve the task challenges in complex scenarios, multi - agent collaboration, which draws on human collective wisdom, has emerged.

In contrast, multi - agent collaboration is like a team of experts, each excelling in their own fields. A multi - agent system is a distributed system composed of multiple agents that can perceive, make decisions, take actions autonomously, and communicate and coordinate with each other. They have their own responsibilities and, through efficient collaboration, jointly produce work results far beyond the capabilities of any single agent.

The advantage of multi - agent collaboration lies in task decomposition and specialization, achieving a more powerful problem - solving ability. Take the digital human anchor as an example. The digital human we see, who can answer questions fluently and has natural expressions, is not powered by a single model but by a collaborative team: a “voice agent” is responsible for generating smooth voices, a “lip - sync agent” ensures that pronunciation is synchronized with lip movements, an “expression agent” controls facial micro - expressions, and a “knowledge agent” is responsible for real - time information retrieval to answer questions. They have their own responsibilities and, through efficient collaboration, jointly present a realistic image far beyond the capabilities of any single agent.

Moreover, a multi - agent system can parallelize the originally linear workflow, significantly shortening the task time. The improvement of its problem - solving ability does not come at the cost of sacrificing efficiency. For example, in software development, one agent can be responsible for writing code, another can conduct testing and find bugs simultaneously, and a third can start writing documentation. A paper from Anthropic shows that a multi - agent system led by Claude Opus with multiple Claude Sonnet agents as subordinates has a 90.2% higher performance than the strongest single agent, Claude Opus 4, with little difference in generation time.

Multi - agent collaboration also brings better fault - tolerance and scalability. A single agent solves problems in a linear process, like putting all eggs in one basket. Once it collapses, has serious hallucinations, or is attacked, the entire task fails completely. In contrast, the team - based multi - agent collaboration naturally has redundancy. If an agent malfunctions, other members can take over part of its work, ensuring that the system does not completely paralyze and is more robust. This distributed architecture also makes it easy to expand the system. When new functions are needed, simply add new expert agents to the team.

If a single agent is a super - individual, multi - agents are more like a collaborative ecosystem. However, everything has two sides, and there are hidden risks in the advantages: the more team members there are, the more complex the coordination becomes. How to make these experts work in unison instead of going their own ways has become the biggest challenge.

Paradox: More Experts, More Troubles?

The more attractive the advantages of multi - agent collaboration are, the more difficult the potential problems are. A paper titled “Why Do Multi - Agent LLM Systems Fail?” reveals the underlying logic of the “more experts, more troubles” phenomenon in multi - agents through in - depth analysis of 7 mainstream MAS frameworks and over 200 tasks: the more a task is split, the more difficult it is to coordinate the consistency of goals, and the more difficult it is to control the output results.

The most obvious problem is the decline in the accuracy rate of some complex problems. Theoretically, more people mean more strength, but the more agents there are, the more difficulties there are in coordination such as communication and monitoring. Misinterpretation or loss of key details can cause sub - agents to act blindly, resulting in a decrease in efficiency. Research shows that agents may misinterpret, modify, or ignore requirements on their own. In the worst - case scenario, the accuracy rate is only 25%, which is lower than the best sampling of a single agent. Take the digital human as an example. If the delay between the lip - sync agent and the voice agent is not perfectly synchronized, the result will be the uncanny valley effect where the voice and lip movements do not match. When real - time information conflicts with the preset script, the digital anchor may seem schizophrenic during the live broadcast and make self - contradictory remarks.

The high communication cost increases the computing power consumption. Agents need to communicate to coordinate, but excessive or inaccurate communication not only incurs high token costs but may also introduce errors and noise. Research on the ECON framework points out that traditional multi - agent debate (MAD) relies on multiple rounds of explicit message passing, and multiple agents may do repetitive work, wasting computing power and possibly producing contradictory results. Data shows that the tokens consumed by agent interaction are about 4 times that of ordinary chats, and for multi - agent systems, it is as high as 15 times. This means that the essence of multi - agent collaboration is still to achieve results by consuming a large amount of computing power. However, due to the complexity of communication, this process is uncontrollable, and the results produced by multi - agent collaboration may not meet expectations.

In addition to the decline in the accuracy rate of some problems and the increase in cost, the division of responsibilities in multi - agent collaboration also hides potential security vulnerabilities. In a single - agent system, if there is a mistake, it is clearly the agent's fault, and the debugging target is clear. However, in a multi - agent system, the final wrong decision is the result of the interaction of multiple agents, and it is difficult to attribute the responsibility to a single individual. It could be that the coordinator decomposed the task incorrectly, an expert agent had hallucinations, or there was a conflict when integrating the correct results of multiple agents and the arbitration mechanism failed. The ambiguity of responsibility allows hackers to manipulate the entire system by deceiving or infecting a single agent.

In short, multi - agent collaboration has both advantages and disadvantages. It has transformed the problem from “how to make an AI smarter” to “how to manage a smart team”.

So, how can we harness this powerful force to make it both effective and not fall into chaos?

How to Break the Ice in Multi - Agent Collaboration

It is not difficult to see that multi - agent collaboration aims to break through single - point intelligence through collective wisdom. However, the tricky part is that training a high - quality team may not be easier than cultivating a genius. Since geniuses always have their own ideas, when several geniuses gather together, coordination and control become difficult.

Since there are so many difficulties, why still take this path?

Because the ceiling is higher.

The limitation of a single agent is the problem of the ceiling of basic capabilities, which can only be solved by scaling the model. In contrast, the errors in multi - agents are engineering and organizational problems, which can be managed and debugged through better system design.

The academic and industrial circles can, through sophisticated system design, control the loss of accuracy rate caused by multi - agent collaboration within a small range, thereby obtaining the huge performance gains it brings in high - complexity tasks, making the multi - agent team both smart and controllable.

To solve the problem of agents acting independently, the system adds a coordinator agent to oversee the overall situation, assign tasks to other agents, and arbitrate conflicts when necessary. For example, Anthropic adopted a “lead researcher - sub - agent” architecture in its multi - agent research system: a lead agent formulates a research plan, then creates multiple sub - agents in parallel to perform different search tasks, and finally the lead agent summarizes the results. This master - slave coordination ensures that the team moves towards a common goal and avoids disorderly competition among sub - agents.

To address the communication challenge, technicians can establish standardized communication protocols to reduce the integration complexity. Agents need to exchange information efficiently and reliably. For this purpose, researchers have proposed various communication protocols and interface standards, such as the MCP protocol and the A2A protocol. Through standardized interfaces, different agents can be easily connected, just like modules in different programming languages interacting through APIs. GenFlow 2.0 is compatible with the MCP protocol and can be flexibly integrated into third - party service ecosystems. This lowers the threshold for developing multi - agent applications and promotes modularity and composability. Developers can connect agents with different functions through standard protocols to work collaboratively, just like building blocks.

To address the potential security vulnerabilities in multi - agent collaboration, researchers can develop more powerful automated failure attribution tools, which can quickly diagnose where the system went wrong, just like a team psychologist, and clarify the responsibility of which agent and which step. Technicians can also introduce adversarial training and resilience design to enable the multi - agent system to learn how other nodes can quickly compensate for failures and maintain overall collaboration when some nodes are compromised.

Of course, we need to note that not all tasks are suitable for multi - agent collaboration. For tasks with a single goal and simple process, using a single agent may be more economical and efficient. The value of multi - agent systems is greater in enterprise - level scenarios where tasks are complex, multiple professional knowledge is required, or high fault - tolerance and parallel processing are needed.

Overall, the current technological trend is to find a balance between distributed agent collaboration and centralized management and control. On the one hand, it is necessary to fully leverage the advantages of distributed decision - making of multiple agents; on the other hand, use coordinators, protocols, and governance rules to constrain and guide the behavior of multiple agents. Only when the technology matures, and the reliability and security are gradually improved, will multi - agent collaboration become more and more effective.

This article is from the WeChat official account “Brain Intelligence” (ID: unity007), author: Shan Hu. It is published by 36Kr with authorization.