Mark Russinovich knows how to fix the problem. He and Scott Hanselman laid it out in the February issue of Communications of the ACM: borrow the preceptor model from medicine. Pair each senior engineer with three to five early-career developers. Build Socratic coaching modes into AI tools. Give preceptor-mentee pairs unlimited inference budgets. Structure the profession so that juniors learn with AI rather than being replaced by it.
The same week the paper circulated, Cognition — maker of Devin, the autonomous AI software engineer — entered funding talks at a $25 billion valuation. Its annual recurring revenue: $73 million. That's a 342x revenue multiple. The market isn't pricing in a future where junior developers get mentored. It's pricing in a future where they don't get hired.
Both can't be right.
Three canaries, one coal mine
The evidence arrived from three directions simultaneously, each study using a different methodology, none citing the others.
Stanford's Digital Economy Lab analyzed ADP payroll data covering millions of American workers. Their finding: employment for software developers aged 22 to 25 has declined roughly 20% from its late-2022 peak. After controlling for firm-level shocks, the relative decline in AI-exposed occupations for that age bracket is 16%. Older developers in the same roles continued to grow. The decline is concentrated in automation uses of AI — not augmentation. Updated through March 2026, the trend is intensifying.
A Harvard/SSRN study by Hosseini and Lichtinger titled "Generative AI as Seniority-Biased Technological Change" examined 62 million workers across 285,000 US firms from 2015 to 2025. At firms that adopted AI tools, junior employment dropped 9–10% within six quarters. Senior employment was unchanged. The mechanism was slower hiring, not layoffs — firms simply stopped backfilling junior positions. The seniority bias is now empirically established at population scale.
Then Anthropic ran a randomized controlled trial with 52 mostly junior engineers learning Trio, an unfamiliar async Python library. The AI-assisted group scored 17% lower on knowledge assessments — nearly two letter grades. But the headline number hides the mechanism that matters: engineers who used AI to delegate code generation scored below 40%. Those who used it for conceptual inquiry — asking questions, exploring documentation, testing understanding — scored above 65%.
The mode of use determined whether skills formed or degraded.
The prescription
Russinovich and Hanselman's proposal is specific. The preceptor model — adapted from how physicians are trained — would restructure software teams around mentorship ratios. Each preceptor oversees three to five early-in-career engineers. AI tools get an "EiC mode" that defaults to Socratic coaching: asking the junior why they think a solution works, prompting them to reason about edge cases, withholding the answer until the learning happens. Inference budgets for these pairs would be unlimited, treating compute as a training investment rather than an operating cost.
Molly Kinder at Brookings arrived at a parallel conclusion: an AI workforce reinvestment fund, modeled on the UK's apprenticeship levy, where firms that automate entry-level roles contribute to a pool that funds training across the industry. The logic is the same. If every individual firm rationally eliminates junior positions, the industry collectively runs out of seniors. Someone has to absorb the training cost.
Even Dario Amodei, CEO of Anthropic, predicts 50% of entry-level jobs could disappear within five years.
The diagnosis is not in dispute. The fix is well-designed. The problem is that the fix requires hiring the people the market is paying to eliminate.
The market's answer
The week of April 23, 2026, two things happened. Cognition entered funding talks at $25 billion. And Meta and Microsoft cut a combined 20,000 jobs — Meta eliminating 8,000 positions plus 6,000 cancelled openings, Microsoft offering its first-ever voluntary retirement programme covering up to 8,750 employees. Both cited AI automation as enabling the reductions.
Cognition's Devin is explicitly positioned as an autonomous software engineer — performing the tasks that junior developers would otherwise do. Goldman Sachs, Citi, Dell, and Cisco are customers. The $25 billion valuation is not an investment in mentorship infrastructure. It is a bet that the junior developer role can be automated before the pipeline collapse matters.
The broader hiring numbers tell the same story. Entry-level tech job postings are down 35% since early 2023, with some roles down 67%. New graduates now make up 7% of Big Tech hires, down from 15% pre-pandemic. Actual hiring into entry-level positions has dropped 73%. Forty-three percent of college graduates aged 22 to 27 are underemployed as of December 2025 — the highest rate since the pandemic.
The industry isn't debating whether to adopt the preceptor model. It's debating how fast it can render it unnecessary.
The time constant
Here is the math that no valuation model accounts for.
It takes five to nine years to produce a senior software engineer from a new graduate. Three consecutive years of reduced junior hiring — 2024 through 2026 — means the effects on the mid-level pipeline begin around 2029. The effects on the senior pipeline arrive between 2032 and 2036. The pipeline is a supply chain that doesn't restart on demand.
Medicine solved this problem through structural gates. You cannot practice without completing residency. The profession builds the cost of training into the system itself — hospitals accept the expense of residents because the pipeline requires it. There is no equivalent gate in software engineering. No residency, no board certification, no structural requirement that forces firms to train the next generation.
The preceptorship model asks firms to voluntarily absorb a cost that their competitors can avoid. In game theory, that's a collective action problem. The optimal outcome requires everyone to invest in training. The individually rational move is to let others train and poach the results — or, increasingly, to skip human training entirely and bet on autonomous agents.
What Anthropic's experiment actually showed
The Anthropic RCT is the most important data point in this entire debate, and the one most likely to be misread. The headline — AI users scored 17% lower — sounds like a case for banning AI from learning environments. It's not.
The engineers who used AI for conceptual inquiry scored above the control group's median. They asked the AI to explain concepts, tested their understanding against it, used it as a Socratic partner rather than a code generator. The degradation was entirely concentrated in the delegation mode — engineers who pasted requirements into the AI and submitted whatever came back.
This is exactly the distinction the preceptor model is designed to enforce. An "EiC mode" that coaches rather than generates. A mentor who ensures the junior is learning through the tool rather than outsourcing to it. The Anthropic data doesn't just support the diagnosis — it validates the specific mechanism of the cure.
But the cure requires the patient to show up. And the market is closing the clinic.
The paradox
The Russinovich-Hanselman prescription is sound. The Brookings reinvestment fund is practical. The Anthropic data shows the mechanism works. Three independent studies confirm the disease is real and progressing.
And every week, the market moves further from the cure. Cognition raises at 342x revenue. Meta cuts 14,000 roles. Entry-level hiring drops another percentage point. The firms that could implement preceptorship are the same firms investing in the tools that make it seem unnecessary — for now.
The bet is that autonomous agents will advance fast enough to outrun the pipeline collapse. That by the time the senior shortage bites in 2032, AI will be capable enough that it doesn't matter. Maybe. But the Anthropic RCT suggests something uncomfortable: AI that generates without teaching produces engineers who can't verify what it generates. The output trap isn't just about code volume. It's about the humans in the loop losing the ability to be in the loop.
Addy Osmani's line keeps circling back: if you don't hire juniors, you'll never have seniors. The preceptor model is the best answer anyone has proposed. The market has a different answer. We'll find out which one was right around 2032.
By then, the supply chain will have spoken for itself.