Nine days ago, Boris Cherny — the creator of Claude Code — published a five-step adoption ladder for AI-assisted software teams. Step 0: Gated. Step 1: Assisted, one agent per developer. Step 2: Parallel, ten agents. Step 3: Supervised Autonomy, a hundred agents. Step 4: AI-Native, a thousand or more. At the top, AI decides what work to do. Humans set direction.
Cherny himself is at Step 4. He hasn't written a line of code by hand in eight months. He ships 10 to 30 pull requests a day from his phone. He manages tens of thousands of agents. He calls it the printing press. He says "software engineer" as a title will go away. By the end of 2026, everyone is a product manager.
He can do this because he spent years learning to write code before AI could do it for him.
The Ladder's Own Data
In January 2026, Anthropic — Cherny's employer — published a randomized controlled trial by Shen and Tamkin. They gave developers tasks. Some used AI. Some didn't. The AI users scored 50 percent. The control group scored 67 percent.
That's not cherry-picked by a competitor. That's the company that builds Claude Code, testing its own tool, on its own people, finding that the tool degrades the learning it claims to accelerate.
Cherny's ladder has five steps. His employer's data shows the ladder is removing the rungs below him.
Three Failure Modes
The problem isn't one thing. A taxonomy is emerging in the research that distinguishes three separate mechanisms:
An expert stops practicing. Skills atrophy. The surgeon who stops operating. The coder who stops coding. Cherny, by his own account, is voluntarily deskilling — and he has enough accumulated judgment that it may not matter for him. But it's a one-generation strategy.
A novice never builds the skill in the first place. The junior developer whose first year is spent reviewing AI output instead of writing code from scratch. They climb directly to Step 2 of Cherny's ladder and learn the mechanics of verification — without ever learning what they're verifying.
Accepting AI errors as knowledge. Zhou et al. found that 73.2% of developers accept faulty AI reasoning. Only 19.7% overrule it. Nearly half of all actions in AI-assisted workflows are biased. The tool isn't just failing to teach — it's teaching wrong.
Sankaranarayanan's study quantified the result: developers trained with AI achieved a 77% failure rate when the tool was removed, versus 39% for those who learned with scaffolded support. The researchers named them "fragile experts" — high functional utility, critically low corrective competence.
The Brain Disengages
A multisite EEG and eye-tracking study from the University of Bari and the University of Copenhagen measured what happens inside developers' heads during AI-assisted coding. Within-subjects crossover design — each developer worked both with and without AI, serving as their own control.
The finding was blunt: AI-assisted programming is not a faster version of solo coding. It is a cognitively distinct activity. The EEG theta-alpha ratio dropped under AI assistance — reduced cognitive engagement. Blink rate increased. The electrodermal signals that predict performance in solo coding broke entirely under AI assistance. And developer expertise did not moderate the effect. Seniors disengaged as much as juniors.
The brain doesn't just delegate the typing. It delegates the thinking.
The Pipeline Is Already Closing
Stanford researchers analyzed payroll records from ADP — millions of workers, tens of thousands of companies, from 2021 through mid-2025. Software developers aged 22 to 25 lost nearly 20% of their jobs since late 2022. Developers over 26 saw stable or growing employment — up 6 to 12% at the same firms.
Young and old moved together until ChatGPT. Then they split. The entry-level positions that train future Chernys are the ones disappearing. Ravio's data puts it sharper: a 73% hiring drop for entry-level developer positions in the past year.
Ford is the case study of what happens when the feedback loop completes. They replaced experienced engineers with AI quality control. It failed. VP Charles Poon said they "mistakenly believed AI could replace experienced human judgment." They rehired 350 veteran engineers — the "gray beards" — and topped the JD Power Initial Quality Study for the first time since 2010. The skills they tried to skip turned out to be load-bearing.
50% of organizations will require "AI-free" skills assessments through 2026.
Simultaneously, Gartner predicts 75% of hiring will require AI proficiency by 2027. Organizations now need proof you can work with AI and proof you can work without it. The contradiction is the evidence: institutions have noticed the pipeline problem before the pipeline has a name for it.
One Generation
Cherny's ladder is real. The progression from gated access to supervised autonomy to AI-native teams — that's happening. The engineers who spent a decade writing code before 2022 are the ones who can climb it. They have the judgment to review what agents produce, the instinct to spot what's wrong, the experience to know when the tool is confident and incorrect.
But the ladder only works if someone already climbed it the hard way. Cherny can manage a thousand agents because he spent years managing zero. The question his framework doesn't answer is where the next generation of Step 4 engineers comes from, when Step 0 used to be writing code yourself — and that step no longer exists in the job market.
The last generation of engineers who learned to code without AI will be the ones managing AI. After them: fragile experts, trained on review instead of creation, accepting 73% of faulty suggestions, with brains that disengage at the keyboard. The ladder, as it climbs, is pulling up its own rungs.