Loop engineering is a pattern where instead of manually prompting AI agents, you design automated systems (loops) that do the prompting. A loop consists of six components: automations (scheduled triggers), worktrees (isolated checkouts per agent), skills (reusable project knowledge), connectors (MCP-based tool integrations), sub-agents (maker-checker split), and persistent memory stored outside the conversation. The maker-checker split is central to safety — a separate checker agent grades the maker's output, ideally using a different model. Key practices include defining explicit exit conditions before running, persisting state to disk or a knowledge graph between runs, and always reading what the loop merges rather than just watching tests pass. Claude Code and Codex both ship all these components today.
