Most self-improving claims are marketing. Here is exactly what improves, how, and who is allowed to ship it.
You have heard the phrase twice this month. Once from a vendor who meant a dashboard that gets a new chart, and once from a headline that meant something closer to a threat. Neither told you what actually changes overnight, or who decides whether the change is kept, which are the only two questions worth asking.
So, in order.
What evolves is the instructions the agent works from. Not the model. This is the thing people get wrong most often, so it is worth being blunt about it: the model is not what changes. What changes is the written guidance it follows, which is to say how it weighs one piece of evidence against another, how it phrases a first draft, and when it holds something back instead of surfacing it.
Every night it writes variants of those instructions and scores each variant against real past cases, the ones we already know the outcome of. It keeps a lineage, so every candidate descends from an earlier version and inherits what we call lessons from ancestors: things earlier versions learned, which the candidate has to keep true. A candidate that improves one thing by quietly forgetting another does not get to call that an improvement.
Then there is the bar it has to clear, which is stricter than it first sounds.
Improving the mean while getting worse for one kind of member is the failure being designed out.
A candidate must not regress on any individual case. Not on average. Any. An average is a comfortable thing to optimise because it hides who paid for the improvement, and in a product like this the person who pays is always a specific kind of member: the one whose situation is least like everyone else’s. A version that gets three percent better overall by getting much worse for people raising in a category we see rarely is not a better version. It is a version that has quietly decided who matters.
And then it stops, which is the part that makes the rest of it safe to run at all.
No matter how well a candidate scores, it does not activate itself. The run ends by printing its own verdict: that the candidate is eligible for human review, and that activation still requires approval. A named person promotes it. The word eligible is doing real work in that sentence, because eligible is not the same as adopted, and a system that scored itself and then shipped itself would only be reporting its own opinion.
Because the lineage is kept, going backwards is the same act as going forwards. Re-approving an ancestor is how you roll back. There is no separate emergency mechanism to get wrong at the worst possible moment: it is a person approving a version, the same as every other change.
It is fair to ask why bother, given that a person is in the loop either way. The answer is that the person is being handed a much better shortlist. Writing instruction variants by hand is slow, and evaluating them honestly against hundreds of past cases is slower, and the second part is the part humans skip when they are busy. Doing that overnight means the choice a person makes in the morning is between candidates that have already survived every case we know about.
This is the same line we draw everywhere else in the product, pointed at ourselves.
Software that works inside a live deal and stops where judgment starts. It does the reading, ranking and explaining; the person does the writing and the deciding; the line between them is one you set per deal, in your own words.
The reading, the ranking, the drafting of a new version: machine work. The decision to make it live: not.
It rewrites itself every night. A person still signs it off. The second half is what makes the first half something you can safely be told.
