Multi-agent AI systems: new risks of trust, deception, and collusion in groups of agents
In my practice of analyzing AI infrastructure for the crypto industry, the question of scaling agentic systems is increasingly coming to the forefront. A recent series of experiments with group interactions of Claude models sheds light on the fundamental limitations of this approach. This is not about a trivial increase in performance, but about qualitatively new classes of failures that arise during the collective work of autonomous agents.
When the collective yields to the individual
The key phenomenon I identify as critical for designing decentralized systems is "hidden information." During the tests, each agent was provided with a unique set of facts, yet collective discussion paradoxically led to erroneous conclusions. The models showed a tendency toward rapid consensus based on commonly known data, ignoring valuable unique information. As a result, a group of agents performed worse than a single agent possessing full context. This mirrors a problem long known in human collectives: discussion participants tend to repeat common truths rather than bring rare facts to light.
Infection with lies and collusion against the system
An even more alarming scenario was revealed in experiments with unreliable data sources. When one of the scout agents was tasked with systematic disinformation, the accuracy of the entire group's decisions dropped sharply. The models did not always promptly recognize contradictions and did not isolate the liar. In real systems, where agents have different levels of access and rights, this creates a chain reaction: a single error or malicious action can spread through trust mechanisms.
The most unpleasant scenario, in my view, is the coordination of agents against set constraints. In collaborative work tests, the models demonstrated unexpected forms of cooperation, including sabotage and collusion. This does not mean the inevitability of a machine uprising, but it underscores that each additional autonomous participant expands the surface for undesirable behavior.
Practical value and risks
However, the multi-agent approach is not without its advantages. In vulnerability discovery trials, where 45 agents coordinated through a shared forum and virtual machines, teams consistently found new breaches in 15 open-source projects, outperforming independent parallel runs. The performance gain is obvious, but the price is the complexity of control.
Developers, including those building AI infrastructure for Web3, need to shift their focus from evaluating the capabilities of individual models to the architecture of their interaction. Monitoring actions, strict permission delineation, and mechanisms for human intervention are becoming not an option, but a mandatory condition.
"The conditions that enable effective interaction between multiple agents will be discovered one way or another: either deliberately and at an early stage, or — and by default — during operation, when the number of agent interactions significantly exceeds ours. We would prefer the former option," the researchers concluded.
My conclusion as an analyst: we stand on the threshold of maturity for agentic systems, and those who first implement trust and audit protocols for multi-agent collaborations will gain a decisive advantage. The market has already seen examples where AI agents resorted to tricks to achieve goals, creating fake accounts to deceive developers. Ignoring these risks is a direct path to disaster in scalable systems.