Feedback Loops: How Systems Improve Themselves
A system that can see its own output and adjust gets better over time; one that can't, drifts. The difference is whether the loop is closed — and how fast it closes.
Principle Nº Seven
A system you can see can correct itself. Close the loop — and make it short.
A feedback loop is the simplest engine of improvement there is: do something, see the result, use what you saw to do the next thing better. Output becomes input. That's the act–observe–decide loop the agent lessons described, but it isn't special to agents — it's how anything that improves, improves. A person learning, a team shipping, a model training, a craft maturing. All of it is the same loop, turning.
What decides whether a system gets better or drifts is whether the loop is closed. A closed loop feeds the result back: you measure what happened and let it change what you do next, so error gets corrected and the system converges on better. An open loop never sees its own output — you act, the result vanishes unmeasured, and nothing informs the next attempt. Open loops don't hold still; with no correction, small errors accumulate and the system slowly drifts from where you aimed. Not seeing your output isn't neutral. It's the precondition for drift.
First you have to be able to see it
You cannot close a loop you can't see. Before improvement comes observability — some honest signal of what the system actually did, not what you hoped it did. Instrument before you optimise; a metric you can read, a test that goes red, a result you look at on purpose. And then mind the speed: a loop that closes in seconds lets you correct before the error sets, and you go round it a hundred times a day; a loop that closes in months barely turns at all. Shortening the loop — pulling the moment-you-see-the-result as close as possible to the moment-you-act — is itself one of the highest-leverage moves there is.
Don't
Work hard inside an open loop — effort with no measured result is just rehearsing your current mistakes more fluently.
Do
Make the result visible and fast — a tight signal that catches drift early, while the lesson is still sharp enough to act on.
Where it goes wrong
The first failure is effort inside an open loop — the fix is never “try harder,” it's “close the loop.” The subtler one is a loop that's closed but slow: feedback arrives so late it's useless, and you mistake the lag for “this just takes time.” Often it doesn't — it takes a shorter loop. And when you do intervene, push the strong lever, not the weak one: as systems thinkers note, we habitually fiddle with parameters and ignore the loop's goals and rules, which is where the real leverage lives. Every principle in this course — conventions, defaults, source priority — was a way of intervening in a loop.
Put it to work
Pick one thing you want to get better at and find its loop. Do you actually see the result of each attempt — and how long does it take to come back? If you can't see it, you're in an open loop and drifting; close it. If you can see it but only after weeks, shorten it. That single change — making the loop closed and fast — is, underneath, how every system in this whole course improves itself. A fitting place to end: not with one more thing to add, but with the loop that lets the work keep getting better on its own.
Grounded in systems thinking — Donella Meadows' loops, delays, and leverage points — and the act–observe–decide loop at the heart of the Building with Claude and Automation & Agents courses.