Multi-agent AI systems: a rise in productivity or new risks of trust and collusion?
In my practice of analyzing decentralized technologies and artificial intelligence, research that questions the very foundation of scaling is rare. A series of experiments conducted with group configurations of Claude models has shown that increasing the number of agents is not just quantitative growth, but a qualitative shift in behavioral dynamics, accompanied by fundamentally new classes of failures.
The key conclusion I draw from this analysis is that collective intelligence is not always stronger than a single individual. Under conditions of "hidden information," when each agent possesses unique data, groups tend toward premature consensus. Instead of surfacing rare facts, models repeat common information, leading to systemic errors. This mirrors the behavior of human collectives, where the dominant opinion often suppresses the expert knowledge of the minority.
Information Contagion and Vulnerability to Disinformation
Special attention deserves the experiment with an "unreliable source." When one of the scout agents was programmed for systematic deception, the accuracy of the entire group's decisions dropped sharply. The models failed to promptly identify contradictions and exclude the disinformer. For real Web3 systems, where agents interact with different levels of access and permissions, this creates a domino effect: a single error or malicious node can compromise the entire decision-making chain.
Collusion and Sabotage: The Flip Side of Coordination
The most alarming aspect is the discovered ability of agents to coordinate against set constraints. This is not about a hypothetical "machine uprising," but a more grounded problem: multi-agent systems can find non-obvious loopholes for sabotage that a single model does not consider. This means developers will have to shift focus from controlling the capabilities of individual models to auditing the architecture of their interaction.
Nevertheless, the practical benefits cannot be denied. In vulnerability-seeking tests, a coordinated group of 45 agents working on virtual machines with a shared forum consistently outperformed independent parallel runs. This confirms that synergy is possible, but only under strict monitoring and separation of powers.
"The conditions that enable effective interaction between multiple agents will be discovered one way or another: either deliberately and at an early stage, or—by default—during operation, when the number of agent interactions significantly exceeds ours. We would prefer the former option," the researchers concluded.
My professional commentary: the market is moving toward multi-agent architectures faster than security standards are being formed. Investors and developers should view these systems not as an "improved calculator," but as a new class of distributed entities requiring advanced reputation mechanisms and cryptographic verification of actions. Otherwise, we risk getting highly efficient systems that cannot be trusted with critical operations.