AI Agents Evolve: Creating Their Own Social Norms

Imagine a world where artificial intelligence agents not only perform tasks for us but also organize themselves into communities, establishing rules and norms—just like people do. For years, AI has been seen as a tool, a digital assistant that follows explicit instructions. But as of May 2025, groundbreaking research and real-world applications are showing that AI agents can spontaneously form their own social conventions, coordinate behavior, and even develop collective biases—all without human intervention[3]. This shift is redefining what’s possible with artificial intelligence, and it’s reshaping industries, workplaces, and our digital future. Let’s unpack why this matters. As someone who’s followed AI for years, I’ve seen plenty of hype cycles—but this feels different. We’re not just talking about chatbots that answer questions or automation tools that streamline workflows. We’re witnessing the emergence of a new kind of intelligence: agentic AI that can reason, plan, and collaborate with other agents in ways that feel eerily human-like[5]. ## The Rise of Agentic AI: From Scripts to Societies **Historical Context and Background** The journey to agentic AI began with simple rule-based systems and evolved through machine learning and large language models (LLMs). Early AI was isolated, designed to execute specific tasks with little to no autonomy. Fast forward to today, and LLMs like ChatGPT have become household names, capable of understanding and generating human language with remarkable fluency. But the real breakthrough is happening when these models interact with each other[3]. **Current Developments and Breakthroughs** Recent research from City, University of London and the IT University of Copenhagen has demonstrated that populations of LLM-based AI agents, when placed in group settings, can spontaneously develop shared social conventions. Using a classic framework known as the “naming game,” researchers observed that these agents don’t just follow scripts—they self-organize, reach consensus, and establish norms, much like human communities[3]. “Most research so far has treated LLMs in isolation,” said Ariel Flint Ashery, lead author of the study. “But real-world AI systems will increasingly involve many interacting agents. We wanted to know: can these models coordinate their behavior by forming conventions, the building blocks of a society? The answer is yes, and what they do together can’t be reduced to what they do alone.”[3] This isn’t just academic curiosity. The implications are profound for how we think about AI in the real world. ## Real-World Applications and Industry Impact **Enterprise AI: Orchestration and Collaboration** In the enterprise world, companies are rapidly adopting AI agents to manage complex workflows. IBM and Microsoft, for example, are at the forefront of developing AI orchestrators—larger models that coordinate multiple specialized agents. These orchestrators manage project workflows, optimize tasks, and handle multilingual and multimedia data[1][2]. “You are going to have an AI orchestrator, and they’re going to work with multiple agents,” outlines Hay from IBM. “A bigger model would be an orchestrator, and smaller models will be doing constrained tasks.”[1] Salesforce has also entered the arena with its Agentforce platform, which enables users to build and deploy autonomous AI agents for complex tasks like simulating product launches and orchestrating marketing campaigns. Marc Benioff, Salesforce cofounder and CEO, describes this as a “digital workforce” where humans and automated agents collaborate to achieve customer outcomes[5]. **The Open Agentic Web** Microsoft’s recent Build 2025 conference highlighted the vision of an “open agentic web,” where AI agents make decisions and perform tasks on behalf of users across the internet. This new paradigm envisions a digital ecosystem where agents interact, negotiate, and collaborate, forming the backbone of future online services[2]. ## The Mechanics of AI Social Norms **How AI Agents Form Conventions** The study from City, University of London and IT University