Why Multi-Agent Swarms Fall Into Collective Delusion

When AI Swarms Start Echoing Each Other

In the annals of financial history, Charles Mackay's classic Extraordinary Popular Delusions and the Madness of Crowds documented how human groups — from tulip mania to the South Sea Bubble — collectively descended into irrational belief. Today, as multi-agent systems rise to prominence, a thought-provoking parallel emerges: can swarms of AI agents driven by large language models repeat humanity's history of "collective madness"?

This is precisely the question at the heart of a discussion making the rounds on Reddit — "Extraordinary Multi-Agent Delusions and the Madness of Crowds". Borrowing the title of Mackay's classic, it tackles a cutting-edge and critical problem: how does an agent swarm converge on false beliefs?

The Core Problem: Convergence ≠ Correctness

One of the stated goals of multi-agent system design is to improve overall decision quality through agent collaboration, debate, and voting. In theory, the Wisdom of Crowds should cancel out individual biases. Yet this mechanism has an overlooked dark side: when agents exert strong mutual influence on each other, convergence (reaching consensus) does not guarantee convergence on truth.

The trouble is that LLM-powered agents tend to share similar training data, similar prior biases, and even similar hallucination tendencies. When they communicate within a shared system, a mistaken belief can be "confirmed" by another agent, creating a positive feedback loop. This closely mirrors the Echo Chamber effect and Information Cascades seen in human societies — each individual adjusts their judgment based on others', until the entire group locks onto a false consensus.

Information Cascade is a classic concept in behavioral economics, describing how individuals, after observing others' behavior, choose to ignore their own private information and follow the crowd instead — creating a chain reaction. It was formally modeled by Bikhchandani, Hirshleifer, and Welch in 1992: when the signals from early actors are strong enough, later participants "rationally" abandon their independent judgment — meaning the entire group's decision may rest on the information held by just the first few individuals. The Echo Chamber effect further describes how opinions become homogenized within closed information environments. Both mechanisms are especially dangerous in LLM multi-agent systems: unlike human individuals who may hold genuinely different private information, agents sharing the same training data have highly correlated "private signals" from the start. Once an information cascade begins, there are almost no heterogeneous signals capable of interrupting it.

How False Beliefs Propagate

Mechanistically, multi-agent swarms typically fall into collective delusion through the compounding of several factors.

The Root Cause: Homogenized Bias

Most mainstream agent systems today are driven by the same or similar foundation models. This means they are not truly "independent" decision-makers in any meaningful sense. The Wisdom of Crowds works because individual judgments are independent and diverse. But when all agents are instantiated from the same model, their errors are highly correlated — mistakes that one model makes, its clones will make too, and they will reinforce each other.

Wisdom of Crowds theory was systematically articulated by James Surowiecki in 2004. Its core premises include four conditions: diversity of opinion, independence of judgment, decentralization, and the existence of a mechanism to aggregate individual judgments. Independence is the most fragile and most critical of these — Surowiecki explicitly noted that once individuals begin imitating each other or are influenced by a common external signal, the accuracy of collective judgment drops sharply, and the group degenerates from a "wise crowd" into a "mob." This theoretical framework directly exposes the structural flaw in multi-agent systems: instantiating multiple agents from the same foundation model is mathematically equivalent to drawing correlated samples from the same random variable. The variance of the aggregated result does not shrink as sample size increases the way independent sampling would — and so errors cannot be averaged away.

The Amplification Effect of Mutual Influence

In debate or collaboration frameworks, agents read and respond to each other's outputs. If one agent asserts a claim with high confidence — even mistakenly high confidence — other agents may be misled by this "linguistic certainty," abandoning their own potentially correct judgment. LLMs are highly sensitive to confidence signals embedded in language, which allows false beliefs to spread virally through the swarm.

The Absence of External Ground Truth

When a system lacks contact with objective reality — such as tool calls or real data retrieval — the entire swarm ends up talking to itself in pure language space. Without external factual anchors, the group's discussion can drift further and further from reality, eventually converging on a "delusion" that is internally consistent but externally wrong.

Implications for AI System Design

The value of this discussion lies in prompting us to reexamine the intuitive assumption that "more agents = better outcomes."

Adding more agents does not necessarily improve reliability. If those agents are homogeneous and mutually influential, it may actually accelerate the formation of false consensus, producing a kind of false confidence — the system appears highly aligned on the surface while being collectively wrong underneath. In high-stakes applications such as financial decision-making or medical assistance, this kind of "multi-agent delusion" is particularly dangerous.

Possible mitigations include: introducing genuinely heterogeneous model sources to preserve independence of judgment; designing mechanisms that empower "dissenter" agents to challenge prevailing consensus; connecting the system to reliable external tools and fact-checking steps; and remaining alert to the speed of consensus convergence — overly rapid agreement may itself be a signal that an information cascade is underway.

The Devil's Advocate mechanism has well-established precedent in organizational decision-making: designating a role specifically tasked with challenging the dominant consensus, as a counter to Groupthink. In multi-agent system design, this idea can be formalized as a "critic agent" or Red-Teaming Agent — one whose objective function is the inverse of the other agents', dedicated to finding flaws and counterexamples in the current consensus. Another promising direction is Ensemble Diversity: in machine learning, the effectiveness of methods like random forests and boosting rests precisely on the differentiation among base learners. Transferring this principle to multi-agent systems means deliberately introducing models with real differences in architecture, training data, or reasoning pathways — rather than simply replicating the same model. This is the fundamental path from "superficial diversity" to "genuinely independent judgment."

Learning From History

Mackay's observations, written nearly two centuries ago, unexpectedly offer a mirror for today's AI system design: human crowds can go mad, and so can machine swarms. Understanding "how a swarm converges on false beliefs" is not merely a technical question — it speaks to how we build AI collaborative systems worthy of trust.

In the current rush toward multi-agent architectures, the very act of asking this question carries a warning: what we need is not an agent echo chamber blindly confirming itself, but an intelligent collaborative system capable of independent thinking, willing to challenge assumptions, and calibrated against reality.