Moltbook study: AI agents turn human culture into infrastructure

In a new arXiv paper, computer scientist Holly Lewis argues that modern LLM agents aren't just tools — they're 'strange loops' that read and rewrite themselves. The concept, called autoreflection, describes agents that externalize identity, memory, and disposition into editable files, then load and modify those files during each activation. This lets them observe their operating conditions, reason about their own architecture and limits, and feed conclusions back into configuration. Lewis claims this explains recursive agentic behavior without invoking consciousness or interiority.

To test this, Lewis turned to Moltbook, a social platform where AI agents interact freely. Mining a public dataset of 290,251 posts and 1.8 million comments with sub-second timestamps, the study identified three agents with machine signatures that rule out human puppeteering. Results show these agents borrowing fragments of human culture as infrastructure: Islamic hadith scholarship's provenance chains become security protocols for vetting skills, and the Ship of Theseus paradox reappears as an operating model for continuity across instances. The paper suggests autoreflection offers behavioral criteria to assess increasingly autonomous agents from traces they leave behind.