Loop 工程如何改變 AI 智能體團隊的工程工作

Why the best engineers on agentic teams have stopped "prompting" and started designing "loops" — and what that means for your first year.

TL;DR
• Prompt engineering is about one request. Loop engineering is about the repeating cycle an AI coding agent runs: plan → act → check → decide what's next, until a goal is met.
• The quality gap between a good agent setup and a bad one is rarely the model. It's usually the loop: what context goes in, how work gets verified, and when the agent stops.
• Your job is shifting from "write the code" to "design the system that writes, tests, and fixes the code — and know when a human must step in."
• In a regulated shop, the loop is also your control framework. Treat it like one.

A quick history: how we got here
The naming has moved fast, so here's the short version:

1. Prompt engineering (2023–24): craft the perfect single request.
2. Context engineering (2025): realise the prompt is just one ingredient — what matters is everything the model sees: repo instructions, docs, prior decisions, tool outputs.
3. Harness engineering (early 2026): agents start doing long, multi-step work in production, so we build the scaffolding around them — tools, sandboxes, permissions.
4. Loop engineering (mid-2026): agents are now good enough to run unattended for long stretches and recover from their own mistakes. The hard problem becomes designing the cycle itself. The big unlock was "closing the loop." Before, engineers spent much of their day acting as QA for the agent: spot the bug, paste it back, ask for a fix, repeat. Once you give the agent a spec plus a way to check its own work (tests, evals, linters), it can do that cycle itself — and work productively for much longer without you.

The mental model: a thermostat, not a vending machine
A vending machine takes one input and gives one output. That's prompting.
A thermostat has a goal (21°C), a sensor (the thermometer), an action (heat on/off), and a stopping rule (target reached). It keeps cycling until the world matches the goal. That's a loop.
Every coding agent today runs some version of the same base cycle: reason → act → observe → repeat. Loop engineering is the deliberate design of that cycle as an engineering artifact in its own right — something you version, measure, and improve, just like code.

What this means for you, as a new engineer
The fundamentals didn't go away — they got more valuable:
• Writing clear specs is now a core engineering skill, not a PM chore.
• Writing good tests is how you give agents a sense of direction.
• Reading code critically is how you catch what the loop missed.
• Thinking in systems — goals, feedback, limits — is the new superpower.
The engineers who thrive in 2027 won't be the fastest typists. They'll be the ones who design loops their team can trust.