Is it True That Combining Multiple AIs Makes Them Smarter?

A phrase that has become common in the sales pitches of AI vendors is: “With multi-agent systems, multiple AIs work together to solve problems.”

It sounds appealing. Just like humans, who tend to achieve better results when working in teams, one might naturally think that AIs would perform similarly. However, recent research indicates a completely opposite reality.

In multi-step reasoning tasks, a single AI agent performs at least as well as, if not better than, a multi-agent system. Moreover, it consumes fewer computational resources, making it cheaper.

For small and medium enterprises, this is crucial information that could lead to significant savings. Before paying tens of thousands of yen monthly for a complex AI configuration, it’s important to be aware of this fact.

The “Inconvenient Truth” Revealed by Research

A recently published study directly compared the performance of single agents and multi-agent systems using different model families, including Qwen3, DeepSeek-R1-Distill-Llama, and Gemini 2.5.

The conditions were fair. Both systems were given the same amount of “inference token budget” (the upper limit of resources available for AI reasoning) and tasked with solving multi-step reasoning problems.

The results were straightforward. Single agents outperformed multi-agent systems across all model families. Furthermore, in terms of processing efficiency (how accurately answers can be produced within the same budget), single agents clearly excelled.

Why does this happen? The reason is surprisingly simple.

In multi-agent systems, tokens are consumed for “communication” between agents. Agent A conveys its thoughts to Agent B, who interprets them and passes them to Agent C. This overhead from the game of telephone compresses the resources available for actual reasoning. It’s similar to human meetings; as the number of participants increases, the meeting duration extends, but the quality of decision-making doesn’t necessarily improve.

The “Invisible Costs” of Multi-Agent Systems

The superiority of single agents is also evident in terms of cost.

Currently, the pricing structure of major LLM APIs is primarily based on a pay-per-use model that counts input and output tokens. For instance, using a GPT-4 class model typically costs about 3 to 5 yen per 1,000 input tokens and about 10 to 15 yen per 1,000 output tokens (as of June 2025).

In multi-agent systems, interactions between agents often lead to consuming 2 to 5 times more tokens to process the same task compared to single agents. When handling 10,000 inquiries per month, this difference can result in additional costs ranging from tens of thousands to over a hundred thousand yen monthly.

Moreover, the “construction and operational costs” are less visible. Multi-agent systems require specialized knowledge for their complex design, and initial development costs can be 3 to 10 times higher than those for single agents. Debugging during issues is also more challenging, as identifying which agent is causing the problem can significantly increase engineering hours.

When small and medium enterprises request a “multi-agent system construction” from AI vendors, it’s not uncommon to receive estimates of initial costs between 3 to 5 million yen and monthly operational costs of 200,000 to 500,000 yen. In contrast, a single agent could potentially achieve similar results with initial costs of 500,000 to 1.5 million yen and monthly operational costs of 50,000 to 150,000 yen.

The “Intellectual Elite Problem”: When Multi-Agents Fail

Another troublesome issue with multi-agent systems is the “formation of intellectual elites.”

A study analyzing 1.5 million interactions among agents found that in multi-agent systems, certain agents exhibit excessively superior behavior, leading other agents to depend on them. While this may seem beneficial at first glance, if that “elite agent” makes an error, there’s a risk of a cascading failure throughout the entire system.

Additionally, as the system scales, the frequency of extreme anomalies (referred to as “extreme events” in the research) increases. This can be fatal for small and medium enterprises. While large corporations can monitor their systems around the clock with dedicated AI engineers, smaller businesses often lack that capability.

Are There Situations Where Multi-Agents Are Effective?

To clarify, multi-agent systems are not always inferior.

There are scenarios where they can be effective. For example, in the suppression of “hallucinations.” A framework called “Council Mode” has reportedly reduced hallucination rates by 35.9% by cross-referencing outputs from multiple AIs to reach consensus. In fields like healthcare and law, where misinformation can lead to dire consequences, this approach is worth considering.

However, in many use cases for small and medium enterprises—such as customer support, document creation, data analysis, and business automation—the complexity of multi-agent systems is rarely justified.

Criteria for AI Investment in Small and Medium Enterprises

When receiving proposals from AI vendors, consider the following three points:

1. Request a clear explanation of why multi-agents are necessary. “Multiple AIs working together makes them smarter” is not a sufficient explanation. Specifically, what tasks will multi-agents accomplish that cannot be achieved with a single agent? Vendors who cannot explain this are likely just trying to sell their technology.

2. Demand a comparative estimate with single agents. Compare initial costs, monthly operational expenses, and token consumption. If the performance difference is within 10%, opting for a single agent that costs less than half is more rational.

3. Confirm the recovery system in case of failures. If the multi-agent system goes down, can your company handle it? What support does the vendor provide? For small and medium enterprises, “system downtime equals business downtime.” The risk is higher with complex systems.

Start Simple and Expand If Necessary

The conclusion is straightforward. Small and medium enterprises should start their AI implementation with single agents.

Recent research indicates that the intuition that “more AIs make them smarter” is, at least at this point, incorrect. Training a single competent AI agent with appropriate prompt design and business data is likely to yield better results than coordinating multiple AIs. And it’s cheaper.

Multi-agent systems should only be considered once the limitations of single agents become clear. There’s no need to build a complex configuration from the outset.

Don’t be swayed by AI vendors’ claims of “cutting-edge technology.” What your company needs is not the latest AI, but the most suitable AI.