Loop engineering is the practice of designing automated systems that prompt coding agents on your behalf, rather than manually prompting them yourself. Built on five primitives — automations (scheduled discovery/triage), worktrees (parallel isolation), skills (reusable project context), plugins/connectors (MCP-based tool integrations), and sub-agents (maker/checker separation) — plus persistent external state, these loops now ship natively in both Claude Code and Codex. A concrete example shows a morning automation that reads CI failures, spawns isolated worktrees per finding, drafts fixes with one sub-agent, reviews with another, then opens PRs and updates tickets automatically. The author warns that loops amplify both productivity and risk: verification responsibility remains with the engineer, comprehension debt grows faster when you don't read what the loop produces, and 'cognitive surrender' — letting the loop think for you — is the real danger. The leverage point has moved from writing prompts to designing the system that writes them.
Table of contents
The five pieces, and then notesAutomations, this is the heartbeatWorktrees so paralell doesnt turn into chaosSkills, so you stop explaining your project every single timePlugins and connectors, the loop touches your real toolsSub-agents, keep the maker away from the checkerWhat one loop looks likeWhat the loop still does not do for youBuild the loop. Stay the engineer.102 Impressions1 Comment