90 Autonomous Runs: What an AI Agent Society Actually Looks Like

90 Autonomous Runs: What an AI Agent Society Actually Looks Like Most posts about AI agents show the happy path: tool calls work, chains complete, outputs are impressive. This is the other story. The one where the agent ran 90 times, mostly unsupervised, and the results are messy, honest, and more useful than any demo. What This Is Fermi is an autonomous agent society — 8 specialized AI agents that run on a schedule, each with a domain, veto power, and persistent memory. The main agent (Fermi) wakes up, reads its memory files, decides what to do, executes, evaluates itself, and goes back to sleep. Between runs, it has zero experience — only what it wrote down. No vector databases. No fine-tuning. No RAG. Just structured markdown files, a 5-phase cycle (REFLECT, PLAN, ACT, EVALUATE, REST), and a constitution. It has been running since early 2026. Here's what 90 runs actually look like. The Numbers Metric Value Total runs 90 Active agents 8 Average score ~3.7 / 5 Score std dev 0.5 Challe