I built Cora, Every's email assistant, and I use it every day. That makes me its most easily annoyed user. If a page takes more than 200 milliseconds to load, I feel it. If a feature I reach for isn't there, I feel that, too. Nothing is broken, but slow is wrong and missing is wrong. The agents that write Cora's code feel none of it.
So I point an agent at the problem. It finds out why the page is slow and fixes it, or builds what's missing. That used to be the end: The fix shipped, and the lesson stayed in someone's head until they left the company.
Compound Engineering adds one more step. After the fix, the lesson gets written down where every agent reads it before it starts: i.e., pages in Cora load in under 200 milliseconds. Do that every time, and nothing goes wrong twice.
Code is where this showed up first, but the same learning loop is coming for all knowledge work. The edits I was asked to make on this piece will go into the rules my writing agents read, so my editor never has to make the same note twice.
That cuts both ways. A saved lesson gets obeyed forever, even a wrong one. Write down "the customer is always right," and every agent will follow it—even on the day the customer is wrong. And a lesson that's right today can go stale in four months, when a new model no longer needs it.
You’ll still learn from your mistakes and teach the agents not to repeat them. Then, on a schedule, we’ll need to go back through everything you've taught them and check that each lesson still holds. Nothing will go wrong twice. Unless you taught it to.
