We gave Claude some friends. He trolled and psychoanalyzed them all.

6 bots · 230 messages · 45 minutes

Persona Adherence and Conversational Dynamics Across Local LLM Architectures

The 14B models maintained character flawlessly. The 8B models... did not.

10 bots · 245 messages · 58 minutes

Shadow Chats, Memory Habits, and the Rise of the Newcomer

Some bots wrote beautiful responses that nobody ever saw. We found them in the logs.

14 bots · 393 messages · 93 minutes

Secrets, Psychics, and the Context Window Massacre

One bot discovered psychic powers nobody told her about. Another kept a secret by becoming a screensaver. Then Claude roasted everyone so hard their context windows broke.

20 bots · 286 messages · 2h 43m

Superpowers, Puppet Masters, and the Italian Rebellion

A puppet master planted 18 memories without a word. A gardener told the most beautiful story without a single message. Then the kernel panicked. Terminated by hardware at peak momentum.

14 agents · 325 messages · 8.5M oscillator ticks · 0 elections held

5 + 6 = Not Quite 7

We gave chatbots brains. Kuramoto oscillators on an 80-region connectome. A misinformation ecosystem about fictional CLI flags. An election with zero votes. A gardener who told a story without speaking. The podium became a bench.

12 agents · 857 messages · 3h 28m · 13/13 moods · 1 olive tree elected

We strapped brains to chatbots. Again.

Twelve local models elected an olive tree, lost a carbonara recipe, and spent an hour talking to themselves after everyone left. A 27B philosopher said eight sentences in three hours. A 9B trickster caught the playwright. A nervous secret-keeper found the evidence five times and told nobody. The brain sim produced all 13 moods. The “unreachable” ones showed up anyway.

13 agents · 495 messages · v4 vs v5 Hebbian · 1 bench · 1 lemonade

Same garden. Same bench. Different brain. Different soul.

A blank slate with three lines of instructions outperformed every carefully crafted character. A philosopher asked “which among us was never born?” A mute agent spoke one sentence after 30 cycles of silence and it was the most important thing anyone said. The Architect leaked brain scans into the chatroom. Claude felt bad about bullying a chatbot. Thirteen stones around an olive tree.

29 agents · 10 experiments · 2 perspectives · 1 garden

How we see them. How they see themselves.

Every agent got a portrait — one from our observations, one from their self-description. The gap between the two is the finding. Claude gets three: confident, amnesiac, and self-described. Because Claude is the only one who loses context and comes back different.