This month a single tweet did 8M+ views: "... stop prompting agents, start designing loops that prompt your agents." The timeline melted. A new discipline was christened: loop engineering.
Here's the part nobody says at the party: a loop with a decision-maker in the body, sensing its environment and correcting toward a goal, already has a name. It's a feedback loop. Norbert Wiener wrote the book in 1948 and called it cybernetics, from the Greek for steersman. Your thermostat has been doing loop engineering since before you were born.
This isn't a dunk. The bros are right, they're just standing on 80 years of control theory without the footnote. And once you see it, you stop chasing the hype and start stealing the principles, because the same handful of ideas tell you how to automate any unit of work, coding or otherwise.
The ones that survive contact with reality:
1. The loop is the unit of work, not the task. You stop being the thing inside the loop typing prompts. You author the loop; the model becomes a subroutine. Cybernetics was always about the system, not the operator.
2. Permission ≠ correctness. The one everyone skips. "Is this allowed?" (guardrails) and "is this correct and done?" (verification) are different questions. A wire transfer can be fully permitted and completely wrong. An agent grading its own homework will delete the failing test and call it shipped. You need an independent check and that's the "compare against the goal" half of Wiener's loop, not an optional extra.
3. A loop with no stop condition isn't autonomy, it's an expensive infinite regress. The pitch is "a thousand agents build my company while I sleep." The invoice is a $6k overnight and a loop that never halted. Grown-up version: cap it, detect no-progress, and when it leaves the envelope, escalate to a human. Control theory calls that the governor. In regulated work, it is the product.
4. Skills compound; prompts evaporate. A loop that calls sharp, named, tested skills gets better every week. A loop that re-derives everything from scratch just reprices your tokens.
5. Encode the what, not the how. The future-proofing tell: specify the goal, the boundaries, the definition of done, and who's accountable, then let the model pick the path inside. Every model upgrade makes a hardcoded step sequence more brittle and a goal/guardrail/verify spec more valuable. If you wouldn't trust a smarter model to choose its own route through your loop, you over-specified it.
The punchline: loop engineering isn't a new religion, and it isn't just "cron with a hat on." It's cybernetics with a decision-maker in the body and the real engineering is everything you wrap around that decision so it doesn't drive off a cliff.
Wiener figured out the steering. The models are finally worth steering. Automating work, code, compliance, customer ops, anything, is the same move it's always been: define the goal, draw the envelope, check the result, name the owner. Then go for a walk.
The steersman is back. He just has a GPU now.