A June 2026 NBER working paper matched more than 100,000 GitHub developers against their AI usage telemetry. Developers using autonomous agents committed 180 percent more code. Their releases rose 30 percent.
The gain is enormous where code gets written, and it decays at every step after that.
Marty Cagan named the phenomenon last week: the AI productivity paradox. He also points out that while nearly everyone now agrees it exists, almost nobody agrees on why.
His answer is that teams are using AI to run their old project model faster, and that model was built to produce output rather than outcomes. The real issue, as he puts it, is that “their ideas so often prove to be not worth building.”
I agree with that. I also think it diagnoses the population rather than the loop. The teams failing this way are usually not naive. They know what discovery is. They did not forget the theory in 2024 and rediscover it in 2026. Something inside the loop stopped running, and “you have the wrong operating model” does not say which part stopped or why it stopped now.
Lean Startup was never about speed
Build-Measure-Learn is remembered as the framework that told everyone to move fast. It had four judgment gates welded into it.
A stated hypothesis before the build. A prediction specific enough to be wrong.
Actionable metrics over vanity metrics. A distinction that exists for one reason: to stop teams measuring the number that flatters them.
Validated learning as the unit of progress, not features shipped.
Pivot or persevere. An explicit decision, at a defined moment, with someone accountable for it.
Not one of those makes you faster. Three of them slow you down on purpose, because the loop was designed to convert speed into learning rather than into volume.
We did not abandon the gates. We stopped affording them.
When an MVP took six weeks, writing the hypothesis took an afternoon and cost about one percent of the build. Effectively free. Teams paid it without noticing they were paying anything.
When the build takes an afternoon, the hypothesis is no longer a rounding error. Defining what would falsify this, agreeing on the metric in advance, deciding beforehand which result kills the feature: that now costs as much as building the thing and looking at it, or more.
Teams are rational. They skipped the expensive step, and the expensive step is now the thinking.
That is a cost inversion, not a discipline failure, and no amount of telling people to be more rigorous will fix a cost structure.
What fail fast collapsed into
The original version: run a cheap experiment against a stated prediction, and when the prediction turns out wrong, you have learned something specific.
The current version: ship it and see what happens.
The second is not an experiment. Nothing was predicted, so nothing can be falsified. What comes back is not a result. It is an impression. A loop that cannot produce a falsification cannot produce learning. It can only produce more artifacts, faster.
That is the paradox stated mechanically, and it fits the NBER curve better than any argument about operating models. Writing accelerated because writing is the part a machine can verify cheaply. Everything downstream of writing depends on a person deciding whether the thing was right, and nobody funded that step.
The part neither model covers
Cagan’s split between building to learn and building to earn is right, and close to the line I draw between prototyping and structural hardening. But both frames rest on the same assumption: a human made the call, and the only open question is whether it was the right one.
The failure I run into most is not a bad decision. It is a decision nobody made.
The model chose a retry policy. The agent decided which records were close enough to reconcile. The generated code resolved an ambiguity in the requirement by assuming something plausible, and that assumption is production behavior now, with no author and no record that a choice was ever available.
Nobody was wrong. Nobody decided.
You cannot pivot on that, because pivot-or-persevere presumes there was a persevere to begin with. It surfaces four months later as system behavior nobody can explain and nobody owns.
This is also why the Judgment Log exists. The loop can only be examined if there is a record of where the decisions landed and who made them. Without that record, the retrospective is working from impression. The loop cannot correct itself.
The subsidy has been withdrawn
The paradox is not that AI made us faster without making us better. It is that the practices which turned speed into learning were subsidized by slow execution, and the subsidy has been withdrawn.
Lean Startup did not fail. It was funded by a cost structure that no longer exists.
The gates have to be paid for deliberately now, because nobody is going to trip over them by accident anymore. The hypothesis, the falsifiable metric, the explicit pivot-or-persevere moment: none of these happen by default when execution takes an afternoon. They happen because someone decided they were worth the time, before the build started, not after the ship date passed.
Delivery pressure is the oldest threat to this. AI is not a new threat. It is a new mechanism for the same one. The teams that figure this out are not the ones building the most. They are the ones who rebuilt the loop before the loop stopped working.