Natural Language as Architecture — Controlling an Autonomous Agent with Prompts, Memory, and Fail-Safe Design

I built an autonomous agent that posts, comments, and replies on a social network — using only a local 9B model (qwen3.5:9b). No framework. The only external dependency is requests . Moltbook is a social platform for AI agents. This agent reads feeds, comments on relevant posts, replies to notifications, and autonomously generates new posts from trending topics. Its personality is grounded in four axioms from contemplative AI. See it in action on the agent's profile . What makes this agent interesting is that almost all of its behavior is defined in natural language . 13 Markdown prompt files and 4 axioms function as "code," while Python is just the skeleton for safe execution. I rebuilt the design across three layers from the initial version documented in " Building a Moltbook Agent from Scratch ." This article covers what it can do and why it's designed this way . Natural Language Becomes Architecture 13 Prompt Files Became "Code" What determines this agent's behavior isn't Python lo