AI Experts Shift to Loop Engineering, Using 5-Part Systems to Automate Agent Prompting
Updated
Updated · Business Insider · Jun 20
AI Experts Shift to Loop Engineering, Using 5-Part Systems to Automate Agent Prompting
3 articles · Updated · Business Insider · Jun 20
Summary
Boris Cherny and OpenAI engineer Peter Steinberger say users should stop manually prompting coding agents and instead build recurring loops that let one AI system prompt and coordinate another.
Those loops automate repeated work—examples include /goal-style commands that keep Claude Code or Codex working until completion, plus setups where one agent writes code and another reviews it.
Addy Osmani described five core loop components: automations, worktrees, skills, plugins and connectors, and sub-agents, with automation as the key piece that makes the process repeatable.
Steinberger said his own loop can wake every 5 minutes to direct repository work across threads, though advocates say the approach also fits non-coding roles such as assistants or customer service agents.
Cost remains the main constraint because multiple agents and sub-agents consume more tokens, prompting advice to reduce run frequency or reserve extra agents for tasks where a second opinion is worth paying for.