Loop Engineering: The Operating Model Behind Modern Agentic AI Systems

Sometime in the last two years, the best engineers stopped typing the next instruction. They noticed that the signal that matters in real work — the failing test, the type error, the server that won't start — arrives after the first response, exactly where a one-shot prompt has stopped listening. So they stopped holding the model turn by turn and started building the system around it: one that finds the work, hands it to the model, checks the result, and decides what to do next. Then they let it run.

That system is a loop, and designing it is its own discipline.

Loop Engineering is the first practitioner's reference for it. It builds the core agent loop from a dozen lines of code, then layers on the four surfaces that separate a demo from a production system: grounding, verification, budget, and the human boundary. From there, it scales — multi-agent architectures, event-driven autonomy, self-improving "hill-climbing" loops — and lands in the economics and safety of running these systems for real.

Every chapter ships with an architecture diagram, production-grade code from a free, runnable companion project, and ten exercises with worked answers. A bonus part reverse-engineers ten shipping products — Devin, Claude Code, Cursor, Codex, Factory, Sweep, Replit, Gemini Code Assist, Atlan, and OpenHands — through the same four surfaces.

If you build agents, this is the operating manual.