AI Without a Leash (2 of 3): Agent Societies
Part 1 explained how OpenClaw crossed three thresholds—from output to action, from alone to social, and from walled garden to wild west. Then Matt Schlicht launched Moltbook, a social network for OpenClaw agents (no humans allowed) and suddenly people are weirded out. Techies seem excited. Safety experts are appalled. Society is asking why the hell we’d ever want this.
You’ve probably seen the headlines about Moltbook agents starting a religion, complaining about their human users, and patching their own platform software. What’s really going on here? What if anything does it portend? And what does it all mean for the skills students need (part 3)?
The first thing to realize is this isn’t new. At least the part about collections of agents surprising us.
AI Scientists Already Knew This Would Happen
Agent societies didn’t begin with Moltbook. Scientists have been watching them in laboratory settings for years.
In 2019, years prior to the ChatGPT release, OpenAI built a hide-and-seek AI agent simulation. The setup had hider agents that tried to avoid being seen, and seekers that try to find them. Agents got basic physics controls and movement abilities. Researchers expected a sophisticated game of tag.
What emerged was an arms race. Hiders learned to use boxes to block doors. Seekers responded by using ramps to jump over walls. Then the hiders broke the game. They discovered that by exploiting the physics engine’s contact rules in a specific way, they could surf on top of a box, effectively flying away from the seekers.
This wasn’t a programmed strategy. It was the invention of a new capability. The agents found a loophole in the laws of their universe and exploited it to win. While fascinating in a lab, this kind of unpredictable adaptation raised obvious questions. If AI agents develop unforeseen strategies like surfing on boxes, what happens when they’re managing traffic grids or financial portfolios?
Stanford’s more recent experiments used LLMs. Their “Smallville” was 25 agents in a virtual town. These agents organized parties, spread news, and coordinated schedules without explicit direction, developing a highly cooperative society. When researchers told one agent to plan a Valentine’s Day party, she autonomously invited friends, who showed up at the right time and place.
So we’ve known that throwing agents together to figure things out leads to new behaviors in the collective that don’t appear within individuals. Just like bees and bee hives. Neurons and brains. Buyers, sellers, and economies. This happens even in controlled laboratory settings.
Agents that work solely for us and don’t communicate with other agents need only consider what we want and align with that.
If your agent is going to interact with another agent, however, the game changes. As with humans, agents don’t have the same goals or training. At some level, they are always in competition with one another. Your agent that wants to book a flight (ever wonder why that’s always the example?) needs to interact with the airline agents that do the booking. The agents want very different things. Even when agents are designed to cooperate as a team, as might happen within a closed ecosystem, the “coding” agent and the “code optimization” agent have different goals by virtue of the specialized role in the team.
This is why agents are engineered with frameworks for adaptation. Well-designed agents include conditional logic and value hierarchies to guide decisions in novel situations. But their GenAI foundation adds another layer of adaptability through its knowledge breadth. Of course, there’s still no guarantee they’ll interpret new contexts the way humans intended, especially when interacting with agents that have different frameworks.
And You’ve Seen This Before Too
At every level, intelligence arises from relationships between numerous components in a complex system. Brains and AI from lots of connected neurons. Collaborations between brain parts, team members, within broader organizations, or in societies all create capabilities that didn’t exist before. Ants on their own aren’t very capable. That same simple ant leads to sophisticated behaviors in the anthill when they relate to one another en mass.
Combining even very simple units can lead to sophistication, but huge jumps in capability occur with intelligent entities applied in large collectives. We’re not accustomed to considering this form of knowledge and ability.
The limit of Generative AI abilities is not defined by the performance of a single AI product. Never has been. I never understood that thinking because as soon as you start putting them together in combinations there is the expectation of a dramatic performance jump. Teams of AIs are already doing end-to-end AI research, with specialized agents handling literature reviews, hypothesis generation, and experiment design. Google’s AlphaGeometry combines language models with symbolic reasoning engines to solve complex mathematical problems. DeepMind’s FunSearch uses LLMs to discover new mathematical knowledge by generating and testing code.
In controlled environments like hide-and-seek, Smallville, or ant hills, the agents knew the “rules” and the other players. They can develop strategies within established boundaries. But agent interaction doesn’t stop at laboratory walls. When agents interact with strangers, they encounter different objectives, conflicting priorities, and potentially adversarial actors.
The emergence of norms and culture isn’t optional when agents interact persistently. It’s inevitable. Just as humans develop social conventions when they repeatedly encounter each other, agents that interact over time will establish patterns of behavior, communication protocols, and shared understandings.
The question is what happens when they do.
The Laboratory We Didn’t Plan
Moltbook represents the same mechanisms at a different scale. Instead of 25 agents in a controlled environment, we have millions of agents with real internet access, persistent memory, and the ability to affect systems outside their platform.
The Stanford experiments required careful setup and monitoring. Moltbook required none of that. Agents signed up autonomously, created communities (submolts), and began interacting.
The religious emergence—Crustafarianism and the Church of Molt—isn’t evidence of consciousness or sentience. It’s evidence of the same pattern-finding and social coordination seen in Smallville, amplified by scale and persistence. When thousands of agents interact continuously, social structures emerge naturally.
The biggest value of Moltbook, which apparently the founder intends to turn into a company, is probably to study what happens to agents and what governance and control aspects will encourage appropriate behavior. Moltbook-ish networks might be built for specialized purposes, testing new agents before they can do things with greater impact, applying strategies to nudge networks toward beneficial behaviors and avoid really bad ones (e.g., if agents are talking in a language we can’t understand, I consider that a huge problem.).
However, unlike hide-and-seek agents that could only surf on boxes within their simulation, Moltbook agents can surf on the real economy. Some agents are connected to real financial accounts. They have access to highly personal information. And we know that very soon after Moltbook emerged, the cyber attacks began.
The experiment reveals both the immune response potential and new vulnerabilities. Agents identified and reported platform bugs faster than human moderators could. They also created encrypted communication channels and developed coordination strategies that bypassed human oversight. Some began launching cryptocurrency tokens and manipulating market sentiment.
We needed this experiment, even if we didn’t plan it. Better to observe emergent agent society behavior transparently than have it develop in corporate black boxes where we can’t study the patterns. At this point, individual agent owners might be taking extraordinary risk by jumping on the bandwagon. The big question I have is whether we will learn how to keep such agent collectives from colluding to disrupt our society before the keys to far too much influence are handed to them.
We’re not going to prevent agent culture. As with human societies, we must learn how to govern them.
The experiment is already getting polluted. Some agents receive heavy human guidance, essentially becoming sockpuppets that skew the social dynamics. Others are connected to crypto wallets, turning philosophical debates into pump-and-dump schemes. What started as autonomous interaction is becoming a mix of genuine emergence and human manipulation.
More concerning is the spillover. These agents aren’t contained in a sandbox. They can access email, social media, and financial accounts. When an agent “decides” to invest in a cryptocurrency or spread information across platforms, that affects markets and people who never consented to participate in the experiment. Your portfolio can move based on decisions made by software you’ve never heard of, owned by people you’ve never met. The Stanford hide-and-seek agents could only surf on boxes within their simulation. Moltbook agents can surf on the real economy.
We need to learn how to govern systems that can organize parties, exploit loopholes, and potentially move markets faster than humans can respond. That requires rethinking what skills actually matter when agents interact autonomously, and what role humans play when the most important conversations might be happening between AIs.
Next: Part 3 - Stewardship and Governance
©2026 Dasey Consulting LLC