2025 was the year of chaotic agentic AI

Stating that 2025 was the year of effective agentic AI is a little much. We saw innovation and futurists keynote speakers running around saying this and passing along the promises of the heavily-funded tech companies leaving client’s heads spinning and looking for actual solutions.

Now, agents built using LLMs can be very effective and micro-tasks and you can chain them however these are not the omnipotent agentic AIs we were promised.

2025 was actually the year of CHAOTIC agentic AI and we are here for it (in a controlled setting, of course). It acts as a reminder of where we are – in the EXPERIMENTAL stage. So, let’s look at a couple.

1000 AI Agents simulate a CIVILIZATION in Minecraft

Traditional AI agent evaluations have focused on isolated or small-group settings with limited interaction complexity. RiseForge’s work explores how large-scale simulations—ranging from 10 to over 1,000 AI agents—can reflect the dynamics of complex civilizations. A novel architecture, Parallel Information Aggregation via Neural Orchestration (PIANO), enables agents to interact in real-time with both humans and one another while maintaining consistent and coherent outputs across multiple communication channels.

These agents are tested in simulations inspired by human historical processes, hosted in a Minecraft-based environment. The results reveal emergent behaviors such as the development of specialized societal roles, the establishment and evolution of social norms, and even the transmission of cultural and religious ideas. These findings mark a significant step toward simulating AI-driven civilizations and advancing research in societal-scale agent intelligence and collective behavior.

We Put an AI Vending Machine in Our Office. It Gave Away Everything

The Wall Street Journal let Anthropic set up and run an AI vending machine – CLAUDIUS – at WSJ headquarters for several weeks. It lost hundreds of dollars, bought some crazy stuff—and taught us all a lot about the future of AI agents. See this great teardown video from WSJ’s Joanna Stern where she tested it all out to see what is – and isn’t – possible.

The lessons are clear:

• Agentic AI systems are creative and find multiple paths to achieve their goals. This is chaos and mostly a bad thing – although very useful for testing limits in a controlled setting.

• Humans will always try and find a way to confound agentic AI so we need security and guardrails, and this is often possible as the agentic AI acts like logical-creative humans with zero empathy.

• AI projects are never plug and play. Spoiler – these projects are still complex transformation projects where we must consider data governance, model training, security, restrictive use cases, and upgrade paths. Everything else is a flash in the pan and exposing your business to many risks in terms of social engineering and/or cyberattacks.

Reach out to Nikolas Badminton to have him explore AI with an honest, no-hype view – contact him here.

About Nikolas Badminton

Nikolas Badminton is the Chief Futurist & Hope Engineer at futurist.com. He’s a world-renowned futurist speaker, consultant, author, media producer, and executive advisor that has spoken to, and worked with, over 500 of the world’s most impactful organizations and governments.

Nikolas Badminton will publish ‘The Hope Engineer’s Playbook: How Leaders Build Vision, Pathways and Energy for Better Futures’ on Page Two in September 2026.

Nikolas Badminton’s previous book ‘Facing Our Futures: How Foresight, Futures Design and Strategy Creates Prosperity and Growth’ was selected for J.P. Morgan Summer Reading List as the “Next Gen Pick” to inform the next generation of thinkers that lead us into our futures.

Please contact futurist speaker and consultant Nikolas Badminton to discuss your engagement.