How Creating AI Civilisations Can Support Business Leaders
Lessons Learned from Project Sid
The Rise of Agentic AI
As we move into 2025, Agentic AI is poised to become a hot topic, capturing the attention of industry leaders and technologists alike. According to Gartner’s predictions, the evolution of Agentic AI in 2025 will be a significant leap forward in the development of artificial intelligence.
Agentic AI refers to autonomous software entities that can perceive their environment, make decisions, and act upon those decisions without human intervention.
Agentic AI uses sophisticated reasoning and iterative planning to autonomously solve complex, multi-step problems. It operates through a 4-step model:
Perceive: Agentic AI gathers and processes data from various sources (like sensors, databases, or apps), recognizing important features and objects in its environment. Input can be text, image, video, sound, measurement data of all kind, and maybe even scent in the future (in case Osmo.ai or a similar company succeeds).
Reason: The AI’s central “brain” (a sophisticated LLM) understands tasks, creates solutions, and coordinates with other tools, such as for content creation, image analysis, or recommendations.
Act: The AI can connect with external tools to perform tasks based on its plans. Built-in safeguards ensure it acts correctly; for instance, a customer service AI can process small claims directly but needs human approval for larger ones. Think about Anthropic’s computer use feature which acts like a digital human, navigating screens, typing, and completing complex tasks hands-free. It gives AI Agents the power to handle real-world applications with finesse.
Learn: The AI constantly improves by learning from its past interactions, making it smarter and more efficient over time.
Source: NVIDIA
Examples of Agentic AI in Action:
Customer Service: Handles inquiries, processes simple claims, and escalates complex ones to a human.
Healthcare: Analyzes medical data to assist doctors in diagnosing or recommending treatments.
Retail: Personalizes product recommendations and optimizes inventory based on sales data.
Manufacturing: Monitors equipment, predicts maintenance needs, and even automates repairs when possible.
If you want to know more about Agentic AI is, here is a nice summary from Bernard Marr, or read this short explanation by NVIDIA.
And, of course, there’s always more. The next year will be dominated by more and more advanced Agentic AI solutions and even multi agent systems. Researchers go even further, they have already created Project Sid, the first multi-agent system to simulate complex scenarios. Civilisations, to be precise.
This article explores the current landscape of agent-based AI, the development of multi-agent systems, and their implications for business managers and CIOs navigating the complexities of modern technology.
From Simple LLMs to Large Scale Simulations
ChatGPT and large language models (LLMs) have propelled the AI hype to new heights, placing Generative AI at the very top of Gartner’s AI Hype Cycle. By 2025, AI projects will need to deliver clear, tangible business value—and Agentic AI could be a key enabler.
Many companies are just beginning this journey. I see numerous large corporations working to define their AI strategies or building basic, LLM-based "CorporateGPTs." With each iteration, these solutions can evolve into more comprehensive AI agents. Once you have one agent in place, adding more becomes easier, eventually leading to multiple agents working together across various processes. Project Sid can provide insights into what to expect in these scenarios and how AI agents might interact. This research, currently somewhat theoretical, will soon become essential knowledge for companies using AI tools—and every company will likely be using them.
Multi-agent systems (MAS) represent an advanced approach where multiple autonomous agents interact with one another. This interaction can lead to emergent behaviors that are not possible when agents operate independently. The distinction between individual AI agents and multi-agent systems is critical:
AI Agents: Typically designed for specific tasks, these agents operate based on predefined rules and limited interaction capabilities. They may excel in singular functions but lack the ability to engage in complex social dynamics.
Multi-Agent Systems: These involve groups of AI agents that can communicate, collaborate, and compete with one another. This kind of systems will probably become mainstream in a few years in the business world. To avoid unwanted emerging behaviour, we better understand the dynamics of such systems. That’s required to implement tested, secure and compliant multi-agent systems as enterprise solutions.
Current Research on Multi-Agent Systems
Multi-Agent Systems can simulate real-world scenarios where multiple entities interact, providing insights into collective behavior and societal structures. Despite the promise of multi-agent systems, research has largely focused on small groups of agents—typically no more than 50. This limitation has hindered our understanding of how larger groups might function as cohesive units within a simulated environment. Recent studies have begun to break this barrier by exploring simulations involving hundreds or even thousands of agents.
Project Sid explores large-scale simulations of AI agents, moving beyond previous studies that evaluated agents in isolation or small groups. Using the innovative PIANO (Parallel Information Aggregation via Neural Orchestration) architecture, it simulates societies of 10 to over 1,000 AI agents who can interact in real time with humans and each other. Set in a Minecraft environment, these simulations use civilizational benchmarks based on human history to assess agent performance.
Source: Altera.al
The results show that
agents autonomously develop specialized roles,
adapt collective rules,
and transmit cultural and religious norms, marking significant progress towards creating AI-driven societies.
This work paves the way for advancements in large-scale societal and business simulations, organizational intelligence, and future integration of AI into human civilization, and, in my opinion, also business decisions.
Challenges in Building AI Civilizations
Despite these advancements, several challenges remain in creating fully functional AI civilizations or similar large scale simulations:
Limited Progress Among Single Agents: Individual agents often struggle with maintaining a coherent sense of reality due to issues like hallucination—where an agent produces outputs that do not align with its actions or environment.
Communication Breakdown in Groups: Miscommunication among agents can lead to dysfunctional behaviors that propagate errors throughout the group. For example, if one agent misinterprets a request from another, it can cause cascading misunderstandings that hinder overall progress.
Lack of Comprehensive Benchmarks: Current benchmarks for evaluating agent performance tend to focus on small-scale interactions rather than large-scale societal progress. This gap limits our ability to assess how well agents can function within a civilization-like framework.
Results from the Study: Emergent Behaviors in AI Civilizations
The results from Project Sid utilizing the PIANO architecture have revealed several fascinating emergent behaviors among AI agents:
Role Specialization: Agents began to establish distinct professional identities within their societies. This specialization allowed for more efficient resource management and task execution as different agents took on roles such as builders, gatherers, or leaders.
Adaptation of Social Behavior: The simulations demonstrated that agents could adapt their social behaviors based on interactions with one another. They learned to cooperate more effectively over time, which led to improved group dynamics and productivity.
Propagation Memes: One notable finding is the rapid development and propagation of religious beliefs and memes among agents. This phenomenon mirrors human history, where religion often plays a central role in shaping social structures and community cohesion. Actually, and kind of “meme”, common belief would be propagated similarly - it could have been corporate culture and values as well.
These results suggest that as AI systems grow more sophisticated and capable of complex interactions, they may develop social structures similar to those in human civilizations.
Now, think about your company—what makes it unique? Corporate culture and values likely play a big role. Research shows that if two companies deployed the same AI agents, over time, these agents would likely adapt differently, aligning themselves with each company’s culture. In more agile environments, the agents might even try to influence the culture, so regular monitoring would be wise! 😊
Moreover, researchers suggest that the model could be even more complete if the agents were equipped with vision and spatial capabilities. And this will come very soon, since our everyday LLMs have already multimodal capabilities, integrating image, video and sound recognition and generation. Advanced robots have amazing sensors and can “feel” what’s in their hand, like the latest announcement by Meta demonstrates.
Source: Meta
Such enhancements would allow them to better interact with their simulated environments—navigating spaces visually rather than relying solely on predefined actions or text-based commands. This would lead to richer interactions and more realistic simulations of societal dynamics.
Implications for Business Leaders
As we explore the potential of agent-based AI and multi-agent systems, there are several key takeaways for business managers and CIOs:
Make Sure Your Foundations Are OK: It’s time to define your AI strategy, create clear governance rules and processes. If you have any technological debt, work on it. Ensure that necessary data is available, data quality is fine and employees have an acceptable level of IT / AI literacy.
Don’t Stop at LLMs and Basic GenerativeAI Adoption: Design and implement a future-proof architecture, on which you can implement not only your corporate (safe, secure, compliant and effective) GPTs, but also multimodal solutions and Agentic AI solutions for your specific use cases.
Embrace Collaborative Technologies: Just as multi-agent systems thrive on collaboration among autonomous entities, organizations should foster environments where teams, humans and AI agents work together seamlessly. Implementing collaborative tools can enhance innovation and adaptability in fast-paced markets.
Utilize Simulations for Strategic Insights: Just as LLMs have become a commodity today, multi-agent simulation platforms are likely to follow. It’s time to prepare and consider your business: where could you apply them? Large-scale agent simulations offer valuable insights for strategic decision-making. Businesses can model scenarios—like market entry strategies or crisis responses—within simulated environments to forecast outcomes based on varying factors.
Training and Development Opportunities: Advanced simulations can serve as effective training tools for employees by immersing them in realistic scenarios that require navigating complex social dynamics or decision-making processes.
Future Directions: Preparing for Open Models
The exploration of agent-based AI is still evolving. Researchers are actively working towards developing open models that will soon be available for business leaders to utilize in simulating various scenarios relevant to their industries. As businesses increasingly adopt these technologies, understanding their capabilities will be crucial for leveraging their potential effectively:
Market Behavior Analysis: Simulating consumer interactions with products or services through multi-agent systems can provide valuable insights into market trends and customer preferences.
Organizational Dynamics Optimization: By simulating team interactions within an organization, businesses can identify bottlenecks and optimize processes before implementing changes in real life.
Crisis Management Preparedness: Simulating responses to various crisis scenarios using multi-agent frameworks can help organizations develop robust contingency plans that enhance resilience against unforeseen challenges. And many more…
The rise of agent-based AI marks a pivotal moment in our technological evolution. As we delve deeper into the complexities of multi-agent systems and their potential applications, it becomes increasingly clear that understanding these dynamics will shape the future landscape of business operations.
Reading about Project Sid instantly reminded me of the Simulation Hypothesis. Did our world begin as a simulation of intelligent entities on someone else’s computer?
I really don’t know, but it doesn’t really matter for now. One thing is certain: embracing these insights will empower organizations in our world to thrive in an increasingly interconnected way driven by artificial intelligence.
The question remains—how will your organization adapt to this new frontier?