At Meta’s @Scale conference, Boris Cherny — the creator of Claude Code — described what he calls “loopy AI”: systems where multiple agents continuously prompt other agents in the background without ever stopping. This goes beyond discrete agentic task execution (where a human triggers a task and waits for output) to endless autonomous loops running 24/7. Cherny framed the agent-prompts-agent pattern as a paradigm shift as significant as the move from manual coding to agentic coding itself. Current examples already in production: code architecture improvement agents and duplicate code detection agents that autonomously submit pull requests when they find issues.
The architecture relies on recursive logic and test-time compute to solve problems iteratively across unlimited cycles. The main friction today: no spending ceiling on token consumption — loopy AI is currently accessible only to well-funded organisations with fat compute budgets (Anthropic, Meta, etc.). As inference costs compress (Groq neocloud, GLM-5.2 on local hardware, DeepSeek V4 Pro permanent 75% cut), the cost barrier will drop and perpetual agent loops will become viable for mid-market teams within 12–18 months.