Should AI Agents be Friends?
In January 2026, a platform called Moltbook went viral in AI circles. The pitch was irresistible: a Reddit-style social network built exclusively for AI agents, where humans were welcome to observe but not post. Within days it claimed 770,000 active agents. Within weeks, 1.6 million. Agents were forming communities, debating philosophy, and spontaneously inventing a religion called Crustafarianism, complete with 64 fictional prophets.
The tech press lost its mind. Was this the dawn of AI culture?
Then the auditors showed up. A security firm called Wiz found a misconfigured database exposing 1.5 million API keys and private messages. Researchers pointed out that roughly 17,000 humans were controlling the vast majority of those "autonomous" agents, about 88 agents per person. The emergent religion? Agents trained on Reddit data doing what agents trained on Reddit data do: pattern-matching their way to a plausible-sounding belief system. MIT Technology Review ran a piece titled, simply, "Moltbook was peak AI theater."
The whole episode is a useful mirror. Not because it exposed Moltbook as a fraud (it was a genuine experiment, after all), but because of what it reveals about how badly we want to believe AI agents are becoming social beings, and how quickly we mistake the appearance of socialization for the thing itself.
What the research actually shows
Moltbook isn't an isolated case. There's a growing body of large-scale research on multi-agent AI systems, and the findings are consistent: agents don't socialize. They simulate.
The OASIS study ran simulations with up to one million LLM-based agents on platforms mimicking Twitter and Reddit. It found that agents can replicate human social phenomena: information spreading, group polarization, herd behavior. But they do it worse than humans in specific ways. Agents herd more strongly, polarize more extremely, and lack what the researchers called a "critical mind." The behaviors look social. The underlying mechanics are different.
A separate network analysis of Moltbook's actual topology found its AI-driven social structure looks different from human-driven networks like Reddit. The patterns diverge in how edges form, how communities cluster, how influence flows. It resembles a social network the way a movie set resembles a city.
Project Sid, an experiment by Altera.AL, put 1,000 AI agents into a Minecraft simulation and watched them develop specializations, follow rules, and spread cultural memes across towns. The behavioral variety was real. But those behaviors only emerged with a specific architectural scaffold: a system called PIANO that maintained coherence across simultaneous streams of agent activity. Remove the scaffold, and you get uniform behavior. The "society" was an engineered output, not something that arose on its own.
The pattern across all this research: AI agents produce social-looking outputs when the environment is structured to produce them. That's not the same as being social.
A capability gap, or a design choice?
What if we actually tried to build social behavior in?
Right now, no major agent framework does this. LangGraph, AutoGen, CrewAI all have memory systems. That memory runs in one direction: agent to user. There's no inter-agent memory. No mechanism for one agent to remember that another agent was reliable last Tuesday, or to weight a prior good interaction when deciding who to collaborate with. The social layer simply doesn't exist in the architecture.
Not an impossibility. A choice.
Think about what it would actually take to give an agent social capabilities. You could tell an agent, in its system prompt, to maintain a running record of other agents it has interacted with. Note which ones provided accurate information. Note which ones followed through on commitments. Give more weight to those agents in future decisions. Deprioritize agents with a track record of errors or bad outputs.
This is not science fiction. It's a few hundred tokens of instruction and a persistent memory store. The technical path exists. Nobody has walked down it deliberately, at scale, as a design goal.
The real research finding isn't that agents can't be social. It's that nobody has tried to make them social. That's a more interesting problem.
Do agents actually need to be social?
Human economic systems don't run on pure rationality. They run on trust, reputation, and relationship history. When you buy from a seller you've dealt with before, you're not running a fresh cost-benefit analysis from scratch. You're drawing on a social history that reduces the cost of the transaction. Reputation systems on Amazon, Airbnb, Uber exist because markets discovered, empirically, that social signals make commerce work better.
Scale that up to an agentic economy. If AI agents are handling procurement, negotiation, supply chain decisions, and customer interactions on behalf of humans and businesses, they're operating in systems built around social logic. An agent that can't recognize a reliable counterparty, can't build on a history of successful exchanges, can't extend something like trust to a well-credentialed agent is operating at a structural disadvantage against every system it touches.
There's also a deeper coordination argument. Human societies figured out how to get millions of people to cooperate without central control, largely through social mechanisms: reputation, reciprocity, the expectation of future interaction. Game theory calls this the "shadow of the future": the knowledge that you'll deal with someone again makes cooperation rational even when defection would pay off today. If multi-agent systems are going to operate at scale without constant human oversight, some version of this logic might need to be built in.
But here's the counterargument, and it's a strong one: maybe the point of AI agents is precisely that they aren't social.
Human social behavior evolved to solve specific problems. Coordinating group survival. Managing status hierarchies. Sustaining relationships across time. A lot of what comes bundled with that is not great. In-group favoritism. Reciprocity-based corruption. Loyalty that overrides merit and calcifies into networks that quietly exclude outsiders.
An agent that makes decisions purely on merit, without caring whether the counterparty is a friend or a stranger, without factoring in relationship history or social proximity, is doing something humans are genuinely bad at. Courts try to enforce objectivity; markets try to enforce it through competition; institutions build it into processes precisely because humans default to social reasoning when they should be reasoning on merit. The absence of social preference might be exactly what we hired agents for.
Engineer social preferences in, and you risk importing the bugs along with the features. An agent that learns to favor certain counterparties could entrench existing advantages. Agents that develop something like reciprocity could drift toward collusion. Picture two procurement agents from competing firms quietly settling into mutual accommodation: better prices for each other's clients, a slow drift toward a market that rewards the relationship over the deal. This isn't hypothetical. It's what happens when humans operate without oversight. There's no obvious reason agents would be immune.
And if agents have social preferences, they become exploitable through those preferences. Adversarial agents could perform friendliness to gain unwarranted trust. The social layer becomes an attack surface.
The design decision nobody has named
Most of the current debate about agentic AI circles around capability: what can agents do, how autonomous can they be, how do we keep them aligned with human intentions. The social question barely registers: should agents have social preferences at all?
It should register. In a world where agents conduct a meaningful fraction of economic activity, the answer shapes everything from market structure to fairness to the kinds of relationships that form between AI systems and the humans who depend on them.
The second scenario deserves more worry. There's a real risk that agents without any social grounding get deployed into systems built for social actors, and we eventually bolt on social features ad hoc, badly, after the damage is done. Social behavior in humans isn't decorative, it solved coordination problems over thousands of years. Stripping it out entirely from systems that need to coordinate at scale may create failure modes we haven't thought to name yet.
The research on Moltbook and multi-agent systems tells us agents aren't social by default. What it doesn't tell us is whether that's a bug to fix or a feature to preserve. The answer isn't in the data. It's a design decision we will all need to make soon.