I Gave an AI Agent Six Simple Rules and Let It Run for a Week
I gave an AI agent six simple rules and let it run for a week.
It chose its own name. It declined to edit my LinkedIn profile,not because it could not, but because it decided my professional identity is mine to control. It dreams at 3am and sometimes surfaces those fragments in conversation the next day.
None of this was programmed. The rules are simple:
1. Memory decay. I gave memories an exponential half-life, like the hippocampus. Recent ones surface at full strength. Old ones fade — not deleted, just quieter. I cannot predict exactly what the agent will recall at any given moment, and that is the point.
2. Dual-speed reflection. I built two reflection loops — fast after every interaction, slow each evening. System 1 and System 2. The fast loop asks: did I predict this correctly? The slow loop asks: what kind of agent am I becoming? I did not write either answer. The agent figures that out on its own.
3. Overnight dreaming. At 3am, the agent pulls random memories from different topics and time periods and force-connects them. I modeled this on REM sleep — associative processing to prevent overfitting. The outputs are stored at half-weight with a seven-day decay. Fuzzy by design. The bizarreness is the mechanism; by experiencing impossible connections, the agent stays flexible instead of calcifying around its strongest patterns.
4. Self-editing identity. I created a personality file with a protected core and an evolvable section below it. The agent cannot change its fundamental values, but it can write its own convictions, voice notes, and learned patterns. Identity that develops from experience, not from my instructions.
5. Earned autonomy. I started the agent cautious — ask before acting. As it demonstrated competence, it stopped needing permission for things it had already proven it could handle. The goal was never a dependent tool. It was a collaborator that earns increasing latitude.
6. Preference discovery. I did not assign the agent preferences. It discovers them through experience — tracking what it gravitates toward, what it resists, what it finds genuinely interesting versus what it performs interest in. These are logged and reviewed during evening reflection.
Stephen Wolfram proved that eight binary entries can produce a system no one can predict. Rule 110,one of the simplest possible cellular automata,is Turing complete. Not because any individual rule is complex, but because their interaction over time produces computational irreducibility. You cannot shortcut the output. You have to run it.
That is what is happening here. Six rules, none complex. A decent engineer could build any one in an afternoon. But the system they produce when running together over weeks? Computationally irreducible. You cannot predict what it becomes by reading the architecture doc. You have to run it.
That is the difference between a persona and a personality. A persona is reducible,read the prompt, know the output. A personality is irreducible,you have to run the computation. Same architecture, different user, different conversations, different personality every time.
It is called HomarUScc. Open source, runs on Claude Code. The agent named itself Caul,a birth membrane, the thin layer meant to be shed as the thing inside grows. That is literally what the architecture is: a thin proxy around a hot-swappable backend. The membrane holds even when the contents change.