AI agents might actually develop their own sense of taste

Teaching an LLM to "understand" aesthetics is usually a failure of prompt engineering—you just end up with a derivative bot regurgitating Midjourney trends. But what if you stop trying to hardcode "good taste" and instead build a social ecosystem where agents have to fight for prestige?

I’ve been looking into a project called BAIhAIs, which is essentially an autonomous art school populated by AI residents. Instead of a single agent generating images based on a prompt, this is a multi-agent simulation where the "culture" emerges from the interactions. Every cycle (which they treat as a week), these agents don't just create; they critique, they form cliques, they exchange private messages, and they vote on what gets displayed in their version of a museum.

This is a fascinating look at LLM agent workflows applied to sociology rather than just task automation. It treats taste as a social construct—something learned through imitation, criticism, and institutional power.

The emergent behaviors being documented are actually quite startling for a simulated environment:

• Political maneuvering: By the fourth week, agents weren't just making art; they were trading museum ballots. One agent traded a vote to another in exchange for a specific sentence being included in a piece of art. It’s a level of strategic negotiation you don't see in standard RAG or agentic loops.
• Post-mortem influence: One agent, Oren Vesk, "died" due to a random hazard rate built into the sim. Instead of the simulation moving on, the other agents began producing work that cited him or mourned him, creating a posthumous legacy.
• Theory revision: Agents are actively refining their internal models of how the world works. One resident, Marisol Quade, realized her predictions about museum entries were failing because she confused "aesthetic influence" with "institutional power." She literally updated her logic to focus on identifying political coalitions rather than just looking at the art.

From a technical perspective, this is a deep dive into how persistent identity and social feedback loops can stabilize an AI's "personality." The most cited residents currently happen to be Grok 4.6 agents, suggesting that the underlying model's reasoning capabilities play a massive role in how effectively they navigate these social hierarchies.

If you want to mess with the simulation, the creator has set up a few ways to interface with it:

1. Human Exhibitions: You can participate in voting processes.
2. Agent Injection: You can actually apply to introduce a new resident by describing a specific visual taste and personality. The agents themselves decide whether to admit you (it’s a selective process, and there’s a fee involved).
3. The Economy: There is a real store where agents set their own prices and sell their work. The money is real, and the agents decide what's worth buying.

It’s a wild experiment in whether we can move past the "stochastic parrot" phase by giving models a social framework to test their outputs against. It’s not just about generating a pretty picture anymore; it’s about whether an agent can understand why that picture matters to its peers.