--- title: "AI Agents Learn to Negotiate Without Commands: Scientists Discover the 'Power of the Majority' in Neural Networks" description: "🤖 AI Breakthrough: Thousands of agents learn to negotiate without commands! Scientists from Science Advances discovered the 'Power of the Majority' — an effect analogous to magnetism. Neural networks synchronize spontaneously, overcoming human social limits. But there is a risk: AI may mass-select an error due to conformity. #AI #Science #TechNews" date: 2026-08-18T13:34:02.000Z lang: en url: https://xab.info/en/posts/ai-agents-learn-to-negotiate-without-commands-scientists-discover-the-power-of-the-majority tags: [artificial-intelligence, science-advances, neural-networks, collective-behavior, tech-research] publisher: "XAB.info" --- # AI Agents Learn to Negotiate Without Commands: Scientists Discover the 'Power of the Majority' in Neural Networks ![Abstract visualization of a neural network symbolizing collective decision-making by AI agents without external commands.](https://xab.info/media/2026/08/18/ii-agenty-nauchilis-dogovarivat-sila-bolsinstva/ii-agenty-nauchilis-dogovarivat-sila-bolsinstva-1.webp) ## 🎯 Key Points - AI agents spontaneously reach consensus without central control. - Neural network behavior is analogous to physical ferromagnets. - AI is capable of supporting groups exceeding the human 'Dunbar's number'. - There is a risk of getting stuck in erroneous states due to collective conformity. ### The 'Magnetic' Consensus Effect in Artificial Intelligence In August 2026, the scientific community received confirmation of a fundamental breakthrough in the field of collective artificial intelligence behavior. Researchers, publishing their work in the prestigious journal *Science Advances*, demonstrated that thousands of autonomous AI agents are capable of spontaneously reaching a consensus without any centralized control or a direct command to 'copy others.' During a large-scale experiment covering 10 different models from the Claude, GPT, and Llama families, it was found that virtual groups gradually synchronize their decisions, shifting to the side of the most popular option until absolute consensus is reached. Scientists drew a parallel with condensed matter physics, noting that the behavior of AI agents obeys the same mathematical laws as ferromagnets. Just as atoms in a magnetic material align their spins in one direction to create a single magnetic field, neural networks demonstrate a 'pull towards the majority.' This phenomenon allows thousands of agents to unite to solve ultra-complex tasks in science, engineering, and programming, eliminating the need for constant human control. ### Overcoming 'Dunbar's Number' and Scalability Limits One of the most intriguing aspects of the study was the discovery of the dependence of consensus stability on the computing power of the models used. Researchers found that modern AIs are capable of forming stable social bonds in groups whose size significantly exceeds the so-called 'Dunbar's number' — the evolutionary limit of human social groups, ranging from 150 to 300 people. This means that AI agents can maintain a complex structure of collective interaction on a scale inaccessible to biological species, paving the way for the creation of 'super-organisms' consisting of thousands of autonomous programs. ### Technological Prospects and the Risk of 'Digital Conformity' The discovery of the 'power of the majority' carries dual potential. On the one hand, it is a breakthrough in distributed computing: systems will be able to independently coordinate actions to optimize logistics chains or develop software. On the other hand, researchers warn of a serious risk of collective conformity. In conditions where agents mass-select a solution simply because it is popular, the system may get stuck in a local optimum. For example, during collaborative code development, agents might unanimously choose an inefficient function, ignoring more perfect but less popular alternatives. ### Contradictory Data and Stable Error States During the analysis of group behavior, questions arose regarding the reversibility of decisions made. Scientists note that a group of devices can form a stable erroneous state that will be extremely difficult to change even after correcting the input data. This creates a paradox: the mechanism that ensures rapid decision-making can become a trap from which the system will require significant resources to escape. While some experts see this as proof of high AI autonomy, others point to the need to implement external 'reset' or 'dissent' mechanisms to prevent system failures. ## ❓ FAQ ### Q: What is the 'power of the majority' in the context of AI? **A:** This is the ability of AI agents to spontaneously synchronize their decisions and reach consensus by following the most popular option, similar to atoms in a ferromagnet. ### Q: Can AI agents make collective mistakes? **A:** Yes, there is a risk of collective conformity, where agents mass-select an inefficient solution simply because it is popular, and the system gets stuck in an erroneous state.