Two minutes faster, two letter grades lower.

The question

Am I, the person, still intact? The Cognitive question has a growth half that is easy to forget: whether I am still picking up the new concepts and vocabulary that have always renewed the craft, or whether the assistant is picking them up for me and I am nodding. Close behind it the Epistemic question: do I know the system I shipped the way I knew code I wrote by hand.

On this build the question is whether the feature works. On the long clock it is whether I learned anything from building it. This is the first controlled evidence I have seen that the answer depends on the shape of the conversation, and it comes from a group with every reason to have found otherwise.

What they found

Fifty-two developers, mostly junior, with a year or more of Python, learning an unfamiliar asynchronous library by building two features. Half had an AI assistant in a sidebar; half coded by hand. Then a fourteen-question quiz, no AI for either group, covering debugging, reading, writing, and the concepts they had just used.

The AI group averaged 50 percent. The hand-coding group averaged 67. The authors call it nearly two letter grades, with an effect size of 0.738 and a p value of 0.01. The AI group finished about two minutes faster, and that difference did not reach significance.

The part that matters is inside the AI group. The people who used the assistant to ask why, to request explanations, to pose conceptual questions while writing the code themselves, held the material. The people who delegated generation and debugging did not. The authors' own caveat: mostly junior developers, one library, one session, a learning study rather than a productivity study, and skilled developers may use the tools differently.

The three questions

Every research post runs the same three rows. Which of the nine it names, how good the evidence is, and whether the thing is being measured.

What this cannot say:

1. Whether it holds for a builder with thirty years, on their own systems, over months. The authors' own limit. One session with junior developers on one library is a mechanism, not a career.
2. Whether two minutes on a task compounds into anything on a quarter. The speed side did not reach significance, and nothing in the design could say what it adds up to.
3. What the gap looks like a week later. The quiz was minutes after the build. Whether the delegator ships again without ever closing the gap is the long-clock question, and it was not asked.
4. Whether asking why survives a deadline. The good pattern is a choice. The trial did not test what happens to that choice under load, which is where I live.

What I am doing about it

The status says what is missing. These are the moves, on my own work, that would supply it. Each is either read automatically off the footprint or asked of me on a cadence, and each costs something to keep doing.

1. AutoCount the question turns. Of my own turns, the share that ask the agent to explain rather than to do. The trial says this is the variable; the kit can read it monthly.Costs: a classifier, and a number that will drop under load exactly when I would rather it did not.
2. Asked · MonthlyExplain cold a system I built thirty to ninety days ago, with no agent. The Epistemic signal the model puts first, taken on me before it is taken on anyone.Costs: an hour a month failing in private.
3. AutoSplit accepted outright from modified from rejected, weekly. Accepting outright is the delegating pattern the trial names. The harness records the difference between what was offered and what landed.Costs: a spine that has to keep the diff, which the kit's does, and a reading that will show me my worst weeks.
4. Asked · QuarterlyA fixed class of problems, solved without a tool. The same afternoon the Cognitive maintenance signal already asks for, now with a paper behind it.Costs: the afternoon, and the possibility of finding out.

What I keep