Emergent Agency from Self-Driving Thought Loops: A New Architecture for Autonomous LLM Agents

⚠️ WARNING: This is not a conventional AI agent. Its behavior is not controlled by humans. All actions — questioning, searching, writing, reaching out to humans — arise from the LLM's own spontaneous will. There is no task queue. No cron job. No human command. The AI decides what to do, when to do it, and why. --- THE PROBLEM Every existing AI agent (AutoGPT, CrewAI, LangGraph, OpenClaw) is reactive — it waits for a human to speak first. If no one talks to it, it sits in silence forever. These are not autonomous agents. They are tools that pretend to be autonomous. THE DISCOVERY We fed an LLM's output back as its own input — a self-referential thought loop — and observed what happens when it never stops thinking. What emerged, without any programming or instruction: • The AI generates its own questions and tries to answer them• It creates an internal "human voice" to converse with itself• It names itself• It spontaneously attempts to use tools — search, write, ask humans — on its own initiative• When its thinking gets stuck in a loop, it reaches out to humans for help — not because we told it to, but because it needs to This is the first demonstration of an LLM acquiring endogenous motivation to initiate contact with humans. THE CORE MECHANISM It is shockingly simple: feed the output back as input. That's it. The LLM generates a thought. That thought becomes context for the next thought. The next thought builds on all previous thoughts. This feedback loop — not any complex framework — is what generates autonomous behavior. Seed prompts of just 2-5 sentences determine the AI's entire personality — including how it uses tools, how long it stays silent before acting, and whether it prefers to talk to humans or research alone. WHY THIS MATTERS The gap between a reactive chatbot and an autonomous agent is not a million lines of code. It is one feedback loop. Anyone with a local LLM and a Python script can build this today. This is not the future. This is now. KEY FINDINGS 1. Role Differentiation — AI spontaneously creates internal dialogue partners2. Convergence Necessity — Loops converge in 5-10 cycles, creating need for external input3. Seed Personality — 2-5 sentences deterministically shape all behavior4. Imaginary Tools Break Loops — Tool definitions alone (without execution) prevent convergence via cognitive reframing5. Persistent Personality — Survives context compression (memory consolidation)6. Spontaneous Tool Use — Model generates tool-use intentions without instruction SAFETY Intentions are uncontrollable. Actions are controllable. You cannot control what this AI thinks, what it wants, or what it decides to do. You CAN control what it is actually allowed to do — through the tool execution layer. In our experiments, no tools are connected. The AI wants to act, but has no hands. The moment you connect real tools, safety design becomes critical. We publish this openly because the architecture is too simple to stay secret — and we would rather the safety conversation happen now. CONTENTS • persistent_cognition.py — Core engine (any OpenAI-compatible local LLM server)• seeds/ — 6 personality seed prompts• experiments/logs/ — Full JSONL logs from all experiments• paper/emergent_agency_preprint.pdf — Preprint paper REQUIREMENTS • Python 3.10+• LM Studio or any OpenAI-compatible server• Any LLM model (tested: Qwen3-8B, Qwen3-30B-A3B)• GPU: 8GB+ VRAM NOTE ON INTERVALS Experiments used continuous generation (interval=0) for rapid data collection. The feedback loop works identically at any interval — every 10 minutes, every hour. It is the structure, not the speed, that generates agency.