The limits of scaling: why more AI agents do not guarantee results

The artificial intelligence industry has long embraced a simple mantra: bigger means better. More parameters, more data, more computational power. A logical extension of this logic became the idea of multi-agent systems — architectures where dozens or hundreds of language models work together, passing information to each other and collectively solving complex problems. However, new research poses a serious challenge to this approach, showing that uncontrolled scaling of agent systems doesn't simply stop delivering results — it actively harms decision-making quality.

Multi-agent systems based on large language models have generated genuine enthusiasm among developers and investors over the past two years. The idea seemed compelling: if one agent can reason and solve problems, then a team of specialized agents should handle far more complex challenges — analogous to how a well-coordinated team of specialists outperforms a single expert. Ambitious projects for autonomous worker systems, agent orchestrators, and multi-level pipelines were built on this intuition. Now it turns out that the analogy with human teamwork doesn't work as straightforwardly as it seemed.

The essence of the revealed problem lies in the nature of the information that agents exchange. When a language model passes the results of its work to another model, that model incorporates the received data into its own context and passes it along — already in an enriched version. With a small number of agents, this process works fine.

But as the system grows, something unexpected happens: the same facts, formulations, and intermediate conclusions begin circulating through the network again and again, accumulating in the context of each subsequent agent. Researchers describe this phenomenon as information redundancy — a situation where signal drowns in a flood of self-repetitions. An algorithm that receives the same fact in ten different formulations doesn't become ten times more confident — rather, it loses the ability to clearly identify what is truly important.

The problem is compounded by the fact that modern language models lack a semantic deduplication mechanism for context. A model doesn't "know" that information about a specific event has already appeared three times in different text fragments — it processes the entire context uniformly, giving disproportionately large attention to repeatedly occurring elements. As a result, the system begins making decisions distorted in favor of the most frequently mentioned data rather than the most relevant data. Accuracy drops precisely when the system is expected to perform at its maximum — when solving complex, multi-step problems requiring subtle analysis.

For the industry, this discovery has quite concrete practical consequences. It means that the current mainstream development path — adding ever more agents in hopes of achieving linear productivity gains — leads to a dead end. Companies that have invested significant resources in building large-scale multi-agent architectures face the need to rethink the fundamental design principles of their systems. This is not about cosmetic changes but about a paradigm shift: from extensive growth in the number of agents to intensive improvement in the quality of their interaction.

Researchers point to several directions that could help developers overcome this impasse. First — intelligent context filtering: agents should not simply pass on all accumulated information but actively select only what is truly necessary for the next link in the chain. Second — development of fundamentally new protocols for inter-agent communication that embed deduplication and information compression directly into the communication process. This will require serious engineering work and likely new approaches to training the models themselves.

The situation reminds one of the history of scaling neural networks, when researchers long believed that simply increasing network depth would yield proportional quality improvements — until they encountered the vanishing gradient problem. Then the solution came in the form of residual connections and normalization. Today, multi-agent systems face a similar challenge, and solving it will likely prove equally non-obvious. One thing can be stated with certainty: the future of powerful AI systems lies not in mindlessly multiplying the number of agents but in each agent's ability to work with information precisely, selectively, and without unnecessary noise.