Google Writes 75% With AI. Source the 25% Who Approve It.
Snap, Google, and Meta just repriced "writes code" to near zero. The scarce role is the engineer who reviews and ships AI code at scale.
Three earnings cycles in three weeks repriced the "writes code" job. Snap said agents now generate over 65% of its new code on the same April 15 memo that cut 1,000 engineers. A week later Sundar Pichai told Cloud Next that 75% of new code at Google is AI-generated "and approved by engineers." Meta's internal target for 2026 is 65% of engineers using AI for more than 75% of their committed code, with layoffs of ~8,000 starting May 20.
If you are still sourcing on "Software Engineer III" and years of experience, you are sourcing the half of the workflow that just got automated. The scarce profile is the engineer who reviews, merges, and owns AI-generated code at production scale. That signal does not live on LinkedIn. It lives in GitHub pull request review history.
The drafting layer is now commodity. The approval layer is the bottleneck.
Read Pichai's sentence carefully. "75% of all new code at Google is now AI-generated and approved by engineers, up from 50% last fall." The clause doing all the work is "approved by engineers." Google has not specified whether the metric is lines, suggestions, commits, or merged changes. The clearest reading is a workflow shift: AI is the default drafting tool, humans are the governance layer.
The trajectory makes this concrete. In early 2024, AI accounted for roughly 25% of Google's code output. By late 2025 it was 50%. Today, 75%. Pichai's own productivity anchor was a complex code migration that agents and engineers completed together six times faster than engineers alone could have a year earlier.
Meta is even more explicit about the role split. The ~7,000 engineers surviving the May 20 cuts are being routed into three new units: Applied AI Engineering, Agent Transformation Accelerator, and Central Analytics. Internal documents describe these groups as "building AI agents capable of handling coding, research, analytics, and operational tasks that have historically been done by human workers." The new role categories, "AI builder," "AI pod lead," "AI org lead," will not surface in LinkedIn title search for months.
So the question for any technical recruiter staffing Q3 is no longer "who can write this service." It is "who can review a thousand agent PRs a week and not ship a regression."
Why review history beats title history
A developer who writes detailed PR descriptions explaining the reasoning behind their changes, engages thoughtfully in code review discussions, and opens well-structured issues is demonstrating exactly the judgment that now bottlenecks production. Those signals live in GitHub events: review counts, review comments, CODEOWNERS membership, merge authority on consequential repos. GitHub Archive captures every public event on the platform, more than 30 million events per month, and makes them queryable. That is the raw substrate for sourcing engineers who review AI code.
The traditional rubric (years of experience, brand-name employers, language list on a resume) was a proxy for code-writing skill. It is a poor proxy for code-reviewing skill, because review skill compounds with judgment and taste, not seat-time. It also does not show up on LinkedIn at all. A junior at a no-name shop who has 400 reviews merged into PyTorch is a stronger reviewer than a Staff Engineer at a FAANG who pushed a hundred small commits to internal repos.
This is why we built Refolk: you describe the profile in plain English ("engineers who have reviewed and merged at least 50 PRs into Hugging Face Transformers, LangChain, or PyTorch in the last 12 months, based in the US") and get a ranked shortlist, instead of a Boolean string that returns the same 200 resumes everyone else is emailing.
The Refolk-index scarcity check
Run the obvious LinkedIn query, "Staff or Principal Software Engineer, skill: Code Review, United States," and you get about 570 profiles, concentrated in the Bay Area and Austin. That is the explicit pool. The implicit pool, engineers whose review history on GitHub demonstrates the same skill but who never typed "code review" into a LinkedIn skills box, is at least an order of magnitude larger. Every Fortune 500 trying to copy Google's workflow is hunting the explicit 570. The teams that ship in Q3 will be hunting the implicit pool.
Snap is the cleanest poach window on the market
The Snap memo is worth re-reading carefully. The 65% AI code disclosure points to engineering. But several analysts have noted the cuts disproportionately hit product, partnerships, distinguished engineers, and directors, not the rank-and-file engineers actually shipping the 65%. That is unusual and useful.
It means Snap engineers who survived April 15 have publicly demonstrated they can ship at a 65%-AI ratio. It also means the engineers Snap did let go (including ML engineers, data scientists, and a slice of distinguished ICs) have the same operating experience without a non-compete cliff. Both cohorts are sourceable right now. Both will be gone by July.
Snap engineers who shipped at 65% AI are the most concentrated agent-reviewer cohort currently on the open market.
Spiegel told investors the cuts would reduce Snap's annualized cost base by more than $500 million by the second half of 2026. The stock rose 8 to 9% on the day. That signal, that the market rewards "fewer engineers shipping more AI code," is exactly the signal every CFO at every public software company is now studying. The Snap layoffs engineer sourcing window is the test case for how the next twelve are going to go.
What "PR-reviewer-grade" actually looks like on GitHub
Concretely, here is the rubric. None of these are exotic. All of them are queryable.
1. Review-to-commit ratio above 1.0
On GitHub's 12-month activity heatmap, the engineers you want have more code-review events than commit events. That is the inverse of the typical "10x committer" profile, and it is exactly the profile the agent era rewards.
2. CODEOWNERS membership on consequential repos
CODEOWNERS files are public on most large OSS projects. Membership signals merge authority, not just contribution. An engineer who is a CODEOWNER on a directory inside Transformers or PyTorch has been formally entrusted with the approval layer.
3. PR descriptions and review comments that explain reasoning
This is the qualitative filter, and it is the single strongest signal you can find on GitHub. An engineer who writes a five-paragraph review explaining why an agent's PR has a subtle correctness bug is precisely the engineer you want approving the next thousand agent PRs.
4. Demonstrated hands-on AI experience
Engineers contributing to Hugging Face Transformers, LangChain, LlamaIndex, or PyTorch have passed a peer review process more rigorous than most technical interviews. That is a stronger AI signal than any line on a resume.
The GitHub PR review sourcing signal is not new. What is new is that, with junior developer employment for ages 22 to 25 down roughly 20% since 2024 per the Stanford AI Index, and with Meta, Snap, and Google all moving engineers into agent-oversight roles, this is now the only signal that matters for staff-level technical hiring.
The quality counter-signal nobody is pricing in
Here is the part of the story that makes the reviewer profile genuinely scarce, not just fashionable. Independent analyses of AI-assisted codebases have measured 1.7x more bugs and 30 to 41% higher technical debt compared to human-written code. The 65% and 75% headline numbers almost certainly include boilerplate, scaffolding, and generated tests.
Which means the limiting factor on agent-era throughput is not how fast the agent drafts. It is how well the human reviewer catches subtle bugs at volume without burning out. That is a skill, and it is unevenly distributed.
Meta is solving for this by formally titling engineers as "AI Pod Lead." Google is solving for this by factoring AI adoption goals into performance reviews and letting select teams use third-party tools such as Claude Code. Snap is solving for this by letting the market sort survivors from the cohort it cut.
Recruiters who want to compete have to solve for it by sourcing on review behavior. The plain-English version of that query, "Find me US-based engineers with 50+ merged code reviews on AI infrastructure repos in the last year who are not currently at FAANG," is the kind of query Refolk runs across GitHub, LinkedIn, and the open web in one shot. That is the entire pitch: you describe the person, you get the people.
Hiring rubrics will lag title taxonomy by six to twelve months
"AI Pod Lead" does not exist as a LinkedIn filter. Neither does "AI Builder," "Agent Transformation XFN," or "Central Analytics." By the time those titles populate LinkedIn search at scale, the engineers under them will already be off the market, retained or counter-offered into the next two-year cycle.
The recruiters and engineering leaders hiring engineers in the agent era who win the next two quarters are the ones who stop sourcing on title and start sourcing on behavior: PR review velocity, CODEOWNERS membership, contribution to agent-infrastructure repos, and reasoning quality in public code review threads. That is what is on GitHub. That is what is not on LinkedIn. That is the trade.
Alexandr Wang and the Meta Superintelligence Labs $1.5B-package crowd are the extreme tail of this same curve. The median version of the bet, the one you can actually staff a team around, is the engineer with 400 thoughtful reviews merged into a serious repo who has never updated her LinkedIn headline.
Find her before Google does.
FAQ
Why is GitHub PR review history a better signal than LinkedIn title history right now?
Because the workflow has flipped. When 65 to 75% of new code is AI-generated, the bottleneck moves from drafting to approval, and approval skill compounds with review reps, not seat-time. LinkedIn titles like "Software Engineer III" were already a weak proxy for code-writing skill. They are a worse proxy for code-reviewing skill, which is exactly what every Fortune 500 trying to copy Google's workflow now needs. The 12-month GitHub activity heatmap, CODEOWNERS membership, and review comment quality are all public and queryable.
How do I source former Snap engineers without burning the relationship?
Move quickly and lead with specifics. Snap cut roughly 1,000 people on April 15, 2026, and the cohort includes ML engineers, data scientists, distinguished engineers, and directors. The most credible outreach references the specific repos or systems they touched, not generic "I saw you were impacted" language. The window before competitors absorb them is short, measured in weeks. Cross-reference public layoff lists with GitHub activity to find engineers who shipped at a 65% AI code ratio and who are now reachable.
What if a strong reviewer has a private GitHub history?
Then you use the surrounding signal. Co-authorship on public commits, conference talks, blog posts on review practice, OSS issue threads, internal documents that leaked into public repos by accident, and second-degree references from CODEOWNERS lists of repos they likely contributed to. A meaningful share of strong reviewers will have at least some public surface area, even if their day-job repos are private. Refolk's value here is combining the public surface area across GitHub, LinkedIn, and the open web into a single ranked answer.
Does this mean junior engineers are no longer worth hiring?
No, but the path has narrowed. Stanford's AI Index shows employment for software developers ages 22 to 25 is down roughly 20% since 2024, which means the market has already partially repriced the "writes code from spec" profile. The juniors who will still be hired in 2026 are the ones who can demonstrate review-grade judgment early, often through serious OSS contributions to agent and codegen infrastructure. That is a higher bar than "graduated from a top CS program," and it is the one the market is now selecting on.