Boris Cherny, head of Claude Code at Anthropic, sparked a conversation by saying he no longer prompts Claude directly — instead, he writes loops. Addy Osmani formalized this as 'loop engineering': an orchestration pattern where developers design automated agent workflows rather than manually prompting. Both Claude Code and OpenAI Codex now ship six core primitives (automations, worktrees, skills, connectors, sub-agents, and memory) that make loop engineering a first-class practice. A key design principle is separating the agent that writes code from a verifier agent that checks it. Osmani cautions about token costs and 'comprehension debt' — the risk of shipping code you never read — when loops run unattended.
Table of contents
From operating the machine to designing the lineThe same building blocks in both productsThe verifier is the part that earns trustCost, correctness and comprehension debtThe road ahead238 Impressions