Loop Engineering: The New Division of Labor in the Agent Era

A new concept is sweeping the AI development community: Loop Engineering. Coined by Peter, author of OpenClaw, the term describes a paradigm shift from writing prompts to designing automated loops that prompt agents. For the past two years, the dominant model has been Prompt Engineering: humans write prompts, models generate results, humans refine. But as coding agents like Claude Code and Codex mature, agents can now read code, edit files, run tests, and call tools autonomously. The bottleneck shifts from "how to write a good prompt" to "how to design a system where agents operate independently." A Loop is an automated closed-loop workflow: discovering tasks (reading issues, CI failures), injecting context, executing in isolation, verifying through tests or agents, and recording state to determine next steps — continue, submit a PR, or escalate. Google engineering lead Addy Osmani identified five core components: Skills (reusable instruction sets), Context Injection (reading current world state), Sub-agents (focused autonomous subtasks), Connectors (post-task actions like PRs), and State Files (persisting progress). Loop Engineering differs from Context Engineering (what agents see) and Harness Engineering (single task stability). Harness is like a kitchen — chef cooks one dish, task ends. Loop is like a restaurant — designed to run continuously with opening hours, kitchen workflow, menu updates built into the system. For clear, verifiable tasks like code testing, CI fixes, and data processing, loops excel. But for product judgment and complex analysis, loops can amplify errors. The ultimate shift: humans move from executor to designer and arbiter, defining goals, boundaries, verification systems, and escalation rules.