The 22-25 Dev Cohort Didn't Disappear. It Reappeared as "Founding Engineer."
Stanford says junior dev jobs fell 20%. Gusto says "Founding Engineer" is up 390% for new grads. Here's how to source the cohort that left LinkedIn.
Stanford HAI's 2026 AI Index landed on April 13 with a number that got read as a eulogy: software developers ages 22-25 are employed at roughly 20% lower rates than they were in 2024, while devs 30+ at the same companies grew. The takeaway from most coverage was that AI ate the junior class. The takeaway you should actually have, if you're hiring, is that the junior class relocated and almost nobody updated their sourcing pipeline.
Look at Gusto's 2026 New Grad Hiring Report and the picture flips. "Founding Engineer" is up more than 390% as a new-grad title. "CEO" is one of the fastest-growing new-grad titles, which is a sentence that should make you stop and reread it. Small businesses (1-49 employees) are forecast to hire about 974,000 new grads between April and September 2026. The cohort big tech stopped hiring didn't vanish. It rerouted into 3-to-7-person seed companies that have almost no LinkedIn footprint.
What the Stanford number actually means
The mechanism behind the 20% drop is task-shaped, not skill-shaped. Stanford's economy chapter, and Anthropic's enterprise telemetry alongside it, attribute the displacement to the work AI now absorbs: boilerplate, scripted tests, CRUD endpoints, routine bugfixes. Anthropic's monitored environments show AI handling roughly 75% of routine programming tasks. That's the exact slice of the stack a 22-year-old used to grind through as a Meta E3 or a Google L3.
Read it the right way and the surviving 22-25 cohort is disproportionately strong on the work AI is bad at: product sense, infra glue, eval design, weird data, agent orchestration. Which is exactly the JD shape that "Founding Engineer at a 5-person company" demands. The job market didn't kill juniors. It killed one specific kind of junior job, and the candidates who would have taken it moved to companies where the remaining work lives.
Why your LinkedIn searches keep missing them
Open LinkedIn Recruiter, type "Software Engineer" and filter "0-1 years experience" plus a target school list. You will get the residual cohort that landed at companies large enough to have a populated LinkedIn company page. You will miss almost everyone at Memora, Hotwash, L2 Labs, Trace Labs, or Neuralk AI, because those companies have 3 employees on the platform and the founding engineer hasn't updated their headline since graduation.
This is the structural problem with sourcing junior developers in 2026. The candidates are real. They ship code daily. They just aren't legible to the tools built for a market where the modal new grad worked at a 5,000-person employer. The Gusto data is blunt: "Software Engineer" as a share of new-grad hires fell from 6.2% in 2022 to 5.4% in 2026, while "AI Engineer" grew 5x. The titles moved. The platforms didn't.
There's a second loop worth naming. "Recruiter" dropped out of the top 20 new-grad titles entirely. The same labor-market shift compressing junior devs is hollowing out the in-house sourcing function that would normally adapt to find them. Sourcing tools are eating their own buyer.
GitHub tool-fingerprints beat pedigree screening
The most reliable signal for an AI-native new grad isn't where they interned. It's what's in their repo root.
Four files have become standard team-context artifacts across Claude Code, Cursor, Codex CLI, and Gemini CLI:
- CLAUDE.md - project context for Claude Code
- .cursor/rules/ - per-repo Cursor rule files
- AGENTS.md - agent instructions, increasingly cross-tool
- SKILL.md - reusable skill definitions
A 23-year-old whose public repos contain a maintained CLAUDE.md and a .cursor/rules/ directory is telling you, with zero embellishment, that they ship with agents daily. Per the Pragmatic Engineer 2026 AI Tooling Survey, 95% of respondents use AI weekly and 70% juggle 2 to 4 tools simultaneously. Claude Code went from zero to the #1 most-used tool in eight months. JetBrains' 2026 Developer Ecosystem Survey clocks Claude Code adoption jumping from 3% in April 2025 to 18% in January 2026, with 46% of devs with 10+ years of experience picking it over Copilot's 9%.
The candidates worth hiring out of the displaced cohort have these fingerprints scattered across personal projects, hackathon repos, and contributions to small OSS tools. The candidates who slept through the shift do not. This is a cleaner first-pass filter than "graduated from a top-30 CS program," because the cohort that should have interned at Meta never got the offer that would have validated the pedigree filter in the first place.
A maintained CLAUDE.md in a 23-year-old's public repo tells you more than a Meta internship that never happened.
This is the exact friction Refolk was built around. "Find devs with public repos containing CLAUDE.md or .cursor/rules committed in the last 90 days, based in EU, currently at a company under 20 employees" is not a Boolean query you can express in LinkedIn Recruiter. It is one sentence in Refolk, and the shortlist comes back joined across GitHub, LinkedIn, and personal sites.
Read "Founding Engineer" as a bimodal title
Glassdoor will tell you "Founding Engineer" averages around $217K. That number is useless because the market is bimodal and the average sits between two real clusters.
The senior cluster is the $250K-plus staff-level role that Recruiting From Scratch describes, often Series A with a real team underneath. The new-grad cluster is the HN "Who Is Hiring" band: $140K-$160K base for one-engineer founding roles, occasionally stretching to a $150K-$220K + equity range for AI infra-flavored work. Robert Half's 2026 Salary Guide puts starting AI/ML engineers at mainstream tech employers around $134K, with Class of 2026 small-business new-grad starting salaries averaging $65,734. The founding-engineer band sits cleanly between those.
Confusing the two clusters produces bad benchmarks in both directions. You'll either overpay for a 23-year-old at a Glassdoor blended rate, or you'll lowball a Series A founding engineer with a new-grad number and lose them in 48 hours.
Where to read the new-grad band live
The cleanest, most current read is the HN "Who Is Hiring (June 2026)" thread, indexed at hnhiring.com. Representative postings from that month:
- Memora, 5-person, EU remote, NestJS / Postgres / PostGIS stack, Founding Engineer
- Hotwash, solo founder building a fire-department after-action-review platform, $40K ARR, 11 paying customers, raising $1M pre-seed, Employee #1 engineering
- L2 Labs, applied AI lab focused on data integration, partnerships with Cornell and NYU, pre-seed
- Trace Labs, physical-AI training data infra, remote USA, founder with a prior exit
- Neuralk AI, Tabular Foundation Models, hybrid Paris/London, seed
None of these companies will surface in a LinkedIn search filtered for "1-10 employees, hiring engineers." Most of them aren't running a LinkedIn job slot at all. The June 2026 thread (id 48357725) is full of one-engineer founding roles in the $140-160K band, and that thread is the price-discovery mechanism for this market.
A sourcing playbook for the displaced cohort
Here's the operating model that actually works for sourcing entry-level software engineers in the post-Stanford-AI-Index world.
1. Build the candidate side off GitHub, not LinkedIn
Start from public repo activity in the last six to nine months. Filter for AI-native tool fingerprints (CLAUDE.md, .cursor/rules, AGENTS.md, SKILL.md). Layer in commit cadence, language mix, and whether their personal projects look like products or like homework. This is the cohort that built things in college because the internship pipeline closed on them, so their GitHub graphs are often denser than the previous cohort's at the same age.
2. Build the demand side off HN and small-cap job aggregators
Index HN "Who Is Hiring" monthly. Cross-reference with YC's company directory, Wellfound, and the seed-stage trackers. Once you have the company list, source the founders' co-author networks and their employees' GitHub contribution graphs. This is where Refolk's "ask in plain English" model collapses what used to be a two-day spreadsheet job into a single query: describe the company shape and the candidate shape together, get a joined list.
3. Replace the resume screen with a public-work screen
The Stanford mechanism implies that the work AI displaced was exactly the work a traditional take-home tested for. Apply the AES-style "no leetcode" frame and screen on a candidate's existing public work plus a 30-minute design conversation. The 22-25s with maintained agent-context files in their repos will dominate this format. The ones without will self-select out.
4. Move fast on the bimodal pricing
If the role is genuinely new-grad founding engineer, anchor at $140-160K base with meaningful equity (1-3% is the live HN band for Employee #1). Don't reach to Glassdoor's $217K average; you'll burn budget and signal that you don't understand the market. If it's a real Series A founding engineer with five years of experience, that's a different conversation and a different shortlist.
The arbitrage window
Aaron Terrazas, Gusto's economist, framed the split in Fortune in May: large companies are "playing defense," small businesses are "playing offense." The Class of 2026 is the first cohort to go through all of college post-ChatGPT. They have Claude Code muscle memory the 30+ cohort is still building.
TrueUp counted 148,092 tech layoffs by June 1, 2026, a 46% acceleration over 2025's daily pace. The conventional read is that this is bad for juniors. The actual read is that the senior layoff wave is pushing experienced ICs into roles that used to absorb juniors, and the displaced juniors are filling founding-engineer slots that, two years ago, would have gone to a 30-year-old with three startup tours.
Founders sourcing this cohort with Refolk are doing the natural-language equivalent of "give me 23-year-olds who ship like 28-year-olds" and getting it back as a ranked list. Recruiters still typing Boolean strings into LinkedIn Recruiter are paying $129K a year to search a population that structurally excludes the candidates they want.
The 20% number Stanford published is real. It just doesn't mean what most people read it to mean. The cohort exists. It's on GitHub. It's on HN. It answers to "Founding Engineer" now.
FAQ
Is the Stanford AI Index junior developer drop actually about AI, or about the broader hiring slowdown?
Both, but the Stanford methodology controls for the slowdown by comparing 22-25 employment to 30+ employment at the same companies. The 30+ cohort grew while the 22-25 cohort fell roughly 20%. That delta is the AI-specific signal. The mechanism (boilerplate, scripted tests, CRUD, routine bugfixes) maps directly to Anthropic's enterprise telemetry showing AI handles about 75% of routine programming tasks in monitored environments. The slowdown is the macro context; the cohort gap is the AI-specific story.
How do I evaluate a 23-year-old founding engineer candidate without traditional pedigree signals?
Start with their public GitHub. Look for maintained agent-context files (CLAUDE.md, AGENTS.md, .cursor/rules/, SKILL.md), commit density over the last 6-9 months, and whether their personal projects look shipped or abandoned. Replace the leetcode screen with a 30-minute design conversation grounded in one of their repos. The AES "no leetcode" frame, and the broader HN "Who Is Hiring" June 2026 norm, both point this direction. The AI-native new grad cohort will dominate this format. Candidates without those signals will self-select out.
What's a fair comp anchor for new-grad founding engineer roles in 2026?
The live HN June 2026 band is $140K-$160K base for one-engineer founding roles, stretching to roughly $150K-$220K + equity for AI infrastructure flavors. Equity for Employee #1 is typically 1-3% at pre-seed. Robert Half's 2026 Salary Guide anchors starting AI/ML engineers at mainstream tech employers around $134K, so the founding-engineer band sits a notch above that with significant equity upside. Do not benchmark off Glassdoor's blended $217K average; it conflates the new-grad cluster with the Series A staff-level cluster.
Where do I actually find these candidates if they're not on LinkedIn?
GitHub first, HN second, personal sites third. Index the HN "Who Is Hiring" thread monthly (hnhiring.com is the cleanest mirror) to find the seed-stage companies hiring founding engineers, then source their employees' and applicants' GitHub networks. For the candidate side, search public repos for AI-native tool fingerprints and filter by commit recency and company size. This is the workflow Refolk collapses into a single plain-English query joining GitHub, LinkedIn, and the open web, which matters because the cohort you're after is the one your existing tools were never built to surface.