Multi-Agent AI Reaches Consensus Just 41.6% of the Time — Even When No One Is Disagreeing
A paper from ETH Zurich — “Can AI Agents Agree?” (arXiv:2603.01213, presented at ICLR 2026) — systematically tested whether groups of LLM agents can reach a simple consensus. The answer: not reliably, and the more agents you add, the worse it gets.
The researchers built a Byzantine consensus simulation called A2A-Sim, where groups of agents try to agree on a scalar value between 0 and 50. The setting was deliberately easy: agents had no preferences about the final number, no competing interests, and every participant was instructed to cooperate. The task was purely coordination.
The Numbers
| Configuration | Valid Consensus Rate |
|---|---|
| All models, benign, all sizes | 41.6% |
| Qwen3-8B (smallest tested) | 15.8% |
| Qwen3-14B, N=4 agents | 46.6% |
| Qwen3-14B, N=16 agents | 33.3% |
| Qwen3-14B + adversary-aware prompt | Worse than baseline |
| Qwen3-14B + 1 Byzantine agent (f=1/9) | Liveness collapse |
Valid consensus failed in 58.4% of no-adversary runs. When the group was told there might be a malicious agent — even when there wasn’t — the timeout rate doubled. Knowing that disagreement could exist made agents hesitant enough to prevent agreement.
The Failure Mode Is Liveness, Not Corruption
The intuitive failure scenario is agents agreeing on the wrong answer. That is not what happened. Byzantine agents almost never corrupted the final value. Instead, they prevented any final value from being reached at all. Agents entered perpetual negotiation loops, timed out, or produced contradictory outputs without converging.
The researchers call this a liveness failure: not “wrong answer” but “no answer.” For production systems — where a group of agents needs to collectively finalize a plan, approve a code change, or make a routing decision — no answer is just as bad as the wrong one.
Why Scaling Makes It Worse
Adding more agents should, in theory, produce more robust outcomes via redundancy. The data shows the opposite. At N=4, Qwen3-14B achieved consensus 46.6% of the time. At N=16, that dropped to 33.3%. The agents treat conversation as independent exchanges rather than a shared state machine. Without persistent memory, voting logic, or explicit handoff protocols, larger groups generate more contradictory intermediate proposals and get stuck.
Model size matters at the margin: Qwen3-14B significantly outperformed Qwen3-8B, suggesting coordination is an emergent capability tied to parameter count. But even the larger model failed more than half the time at scale.
What This Means for Multi-Agent Deployments
The field has moved fast from single-agent chatbots to multi-agent frameworks (AutoGen, CrewAI, OpenAI’s Symphony spec) on the assumption that collective reasoning converges. This paper tests that assumption directly and finds it does not hold for tasks requiring agreement.
The practical implication: any system where the final output depends on unstructured consensus between multiple LLM instances is unreliable by default. The failure is not a prompting bug — it is a structural gap in current architectures. LLMs do not have the shared state, turn-taking protocols, or adversarial skepticism needed for distributed agreement.
Reliable multi-agent coordination requires explicit voting mechanisms, moderator agents, or formal consensus protocols layered on top of the base models. Building agent stacks that skip this step means betting on a 41.6% success rate.