[ICLR 2026] Can AI Agents Agree? The Fragility of Consensus in LLM Swarms

Executive Summary

As we move toward a world of "Agentic Workflows," we often assume that multiple LLMs can coordinate to solve a problem. However, this paper from ETH Zurich reveals a sobering reality: LLM agents struggle to reach consensus even when they have no reason to disagree.

Through a series of controlled experiments using the Qwen3 family, researchers found that the ability to reach a common decision (Consensus) is not a robust capability but a fragile one. As group size grows or a single malicious "Byzantine" agent enters the mix, the system's ability to "agree to agree" (Liveness) collapses, often resulting in endless loops or timeouts rather than a unified outcome.

Problem & Motivation: The Gap Between Math and Prompting

In classical distributed systems, the Byzantine Generals Problem is a solved mathematical riddle. We have deterministic algorithms (like PBFT) that guarantee agreement as long as fewer than 1/3 of the nodes are malicious.

However, LLMs are not deterministic machines; they are stochastic reasoners driven by natural language prompts. The authors identified a critical gap:

1. Lack of Robustness: We don't know if LLMs can handle "strategic" or "malicious" peers.
2. Scale Issues: Does adding more agents make consensus harder or easier?
3. The Liveness Trap: Do agents fail because they are tricked into the wrong answer (Safety violation), or do they just fail to ever stop talking (Liveness violation)?

Methodology: The Scalar Consensus Game

The researchers built A2A-Sim, a playground where $N$ agents try to agree on a number between 0 and 50.

Figure 1: The interaction loop. Agents propose values, justify them with text, and then vote to either 'stop' (if they think everyone agrees) or 'continue'.

The "Secret" Byzantine Agent

In some runs, the authors introduced agents instructed to disrupt consensus while appearing cooperative. These agents don't just provide random numbers; they use LLM reasoning to identify the most confusing values to propose, effectively "gaslighting" the honest agents into doubting if agreement has been reached.

Experimental Insights

The results were surprisingly pessimistic across several dimensions:

1. The Cost of Paranoia

Even when no malicious agents existed, simply mentioning the possibility of an adversary in the system prompt caused agents to become "hesitant," leading to higher timeout rates. Trust, it seems, is a prerequisite for LLM efficiency.

2. Scalability vs. Reliability

Contrary to the "wisdom of the crowd," larger groups of LLMs performed worse. Figure 2: Success rates drop as group size N increases from 4 to 16.

3. Liveness is the Real Killer

The study found that Byzantine agents rarely "corrupt" the final value (getting honest agents to agree on a "bad" number). Instead, they succeed by preventing any agreement at all. They trap the honest agents in a state of perpetual negotiation, causing a total loss of liveness.

Figure 3: As Byzantine agents (B) increase, the "Valid Consensus" (blue) disappears, replaced by "Timeout" (missing bars).

Critical Analysis & Conclusion

This paper serves as a vital "vibe check" for the multi-agent industry. We are currently building complex agentic stacks (Auto-GPT, OpenDevin, etc.) on the assumption that agents can coordinate. Yet, this research shows that even for a no-stake, simple scalar value, they fail nearly 60% of the time in benign settings.

Key Takeaways for Developers:

• Model Size Matters: Qwen3-14B significantly outperformed 8B, suggesting that the "social reasoning" required for consensus is an emergent property linked to parameter count.
• Don't Over-Prompt for Risk: Warning agents about potential adversaries can paralyze their decision-making.
• Byzantine Robustness is Missing: Current LLMs lack the "skepticism" or "voting logic" needed to ignore a malicious minority.

Limitations: The study only used one model family and a restricted "no-equivocation" threat model. Future research needs to see if specialized "Consensus-tuning" or "Chain-of-Verification" can bridge the gap between AI agents and the rigorous guarantees of classical distributed systems.