I want to write something useful about engineering productivity in 2026. The discourse has been bad in a specific way: there's a "AI makes everyone 10x" camp and a "AI just replaces juniors" camp, and they're arguing past each other because they're both wrong. The interesting thing is what's actually happening to the distribution of engineering output, and what it means for your career if you're somewhere on that curve.

Here's the honest version, with the numbers.

What's actually happening (the numbers)

The productivity gap between top engineers and median engineers has roughly tripled since 2024. That's the load-bearing fact.

GitHub's 2026 Octoverse data shows the median engineer at large tech companies is now merging 11 PRs per week, up from 4 in 2024. That's a 2.75x lift at the median. Real, but not the "10x" the discourse promised.

At the 90th percentile, engineers are merging 38 PRs per week, up from 8 in 2024. That's a 4.75x lift. The top is moving faster than the middle.

At the 99th percentile, the number is 80+ PRs per week. A handful of engineers at Anthropic, Cursor, Vercel, and OpenAI have publicly disclosed PR counts that would have been physically impossible to type in 2024. The work isn't typing anymore; the work is judgment, review, and orchestration.

Cursor's internal data, leaked in a recent Latent Space transcript, puts the gap between their top-quartile and bottom-quartile users at 8.4x measured by "tasks completed per active session." A year ago the same number was 3.1x. The distribution didn't get faster uniformly — it stretched.

Two years ago an engineering team was a roughly normal curve of output. Mid-2026, it's a power law. The top of the curve is pulling away at a rate that's hard to look at calmly if you're in the middle.

Why the gap widened (mechanism matters)

The naive theory was: "AI is a force multiplier — everyone improves by some factor." That theory predicts the distribution shifts right but keeps its shape.

That's not what happened. What happened is that AI tools turn out to amplify a specific skill — call it "decomposition and review" — that has very different baselines across engineers.

Here's the mechanism. The work AI does best is the middle layer: "given a clear specification of what you want, generate a candidate implementation." It does this fast and cheap. The work that still belongs to the human is the layer above ("what should we build, and how do we know if it works") and the layer below ("is this implementation actually correct, and is it the right shape for the surrounding system").

Engineers who were already strong at the above-and-below layers got a massive force multiplier — they can now process 5-10x more candidate implementations per unit time, and their batting average for picking the right one is unchanged. So their output went up roughly linearly with their AI throughput.

Engineers who were weak at the above-and-below layers got a much smaller multiplier. They generate candidates fast, but their batting average for picking the right one didn't improve, and now they're producing bad code 5x faster. After review cycles and rework, their actual shipped output went up maybe 1.5x.

The skill that compounds with AI is taste. Taste was always distributed unevenly. The unevenness is now reflected in shipped output at a magnitude it never was before.

Who's pulling away, and where

Specifics matter. The gap looks different by role:

• Senior backend engineers at agent-heavy companies: the biggest gap. Cursor, Anthropic, and Vercel internal numbers all show 6-8x spread between top and median performers. The work — design APIs, integrate models, evaluate quality — is the work AI tools amplify most.

• Senior frontend engineers: smaller gap, around 3x. Design judgment and user-facing taste don't compound with AI the same way. AI helps with the boilerplate; the hard parts are still human.

• Staff/principal engineers: the gap is enormous but inverted from junior intuition. A great staff engineer in 2026 ships more design docs, more code reviews, and more architectural decisions than they ever have — they're the orchestration layer for a stable of agents. A weak staff engineer in the same seat is competing against agents on raw output, and losing.

• Junior engineers: the most exposed cohort. The traditional "write the boilerplate ticket to learn the codebase" path no longer exists in most companies. Companies that were hiring 100 juniors in 2023 are now hiring 20 and pairing each with 5 AI agents. The ones who get hired are the ones who already operate like junior staff engineers — taste-forward, judgment-forward.

The roles compressing fastest aren't the ones the 2023 discourse predicted. It's not "AI replaces juniors." It's "AI flattens the middle." The middle 60% of engineers — competent, productive, not exceptional — is being squeezed because the top of the distribution can now do their work, faster, with less coordination cost.

What it means if you're a working engineer

If you're an engineer in 2026, here's the realistic picture for the next 2-3 years.

The dichotomy isn't "you're safe" versus "you're replaced." It's "you're compounding" versus "you're being out-shipped by people who are." Three years out, both of those become highly visible in salary and seniority data. The compounding cohort is going to pull away in compensation, scope, and influence in a way that's hard to catch up to once it's locked in.

Compounding: engineers who've internalized AI tools as a leverage layer over their existing skill, who treat code as something they edit and review more than something they type, and who own outcomes (decisions, designs, evaluations) rather than tickets. These engineers are shipping more, going senior faster, and earning meaningfully more — internal Cursor + Anthropic data suggests something like a 50-70% comp premium for engineers in the top quartile of "AI fluency" within their level.

Out-shipped: engineers who use AI tools as autocomplete on top of their existing workflow. They got a 1.5x lift. They feel productive — autocomplete works! — but the goalposts moved 4x for the same level. Two years from now, the rubric they're being measured against will be calibrated to the top quartile, and "decent at my job" will read as "well below the new bar."

The gap isn't moral — it's structural. The people falling behind aren't worse engineers than they were in 2023. The environment changed faster than their muscle memory.

What to actually do

I'm going to be specific. Three moves, and they compound — same pattern as the AI/jobs piece, because it's the same underlying mechanic.

Move one: ship more candidates and review more candidates

The throughput question is real. Working engineers in 2026 are processing 5-10x more "generated implementations to review" per unit time than in 2024. If you're still writing each line, you're at a 5x disadvantage on volume alone.

Pick one AI tool — Claude Code, Cursor, Codex, your call — and use it for your actual work, not for tutorials. The right benchmark is: "in this PR I just shipped, what percentage of the code did I type versus accept-with-edits?" If that number is over 70% in 2026, you're operating with 2024 muscle memory. The target is somewhere between 20 and 40% for most engineers.

Move two: move your time toward the layers AI is bad at

The layers AI is bad at (in 2026, this changes every six months — re-check often):

• Picking the right thing to build, given fuzzy product context
• Knowing whether a generated implementation is actually correct for the surrounding system
• Tradeoff calls when two valid options have different long-term costs
• Operational judgment: deploy strategy, monitoring, incident response
• Working through ambiguity with humans — design reviews, stakeholder negotiation, mentorship

If your day is mostly "implement this well-specified ticket," your time allocation is at the layer AI eats first. Get more of the layers above. The most actionable version: take on the design work that your manager or staff engineer is happy to delegate. Own outcomes, not tickets.

Move three: review your own AI output ruthlessly

The trap of AI-assisted engineering is that it generates code that looks plausible at a glance. Plausible-looking code has a much higher bug density than code you wrote yourself, because you didn't think through the edge cases as you typed.

Develop a habit: every PR you ship with AI assistance, you re-read line-by-line and ask "do I understand why every line is here, and would I write it the same way?" If the answer is no, edit. If the answer is "I don't know, but it works," that's the muscle that atrophies fastest. The engineers in the compounding cohort are the ones who use AI as a leverage layer over deep code-reading discipline, not as a substitute for it.

The 10x engineers in 2026 aren't typing 10x more code. They're reviewing 10x more code, and they're approving the right 30% of it.

What I think the next three years actually look like

A continued widening of the distribution, then a partial collapse as the org charts catch up.

• Through 2027: the gap keeps widening. Top-quartile engineers compound. Bottom-quartile flatten. Comp distributions stretch — the comp premium for top-quartile AI-fluent engineers grows past 100% at the same level.

• 2028-2029: orgs restructure around the new distribution. Fewer engineers per team. Each engineer responsible for more surface area. The "team of 5 engineers" gets replaced by "team of 2 senior engineers and an agent fleet" in a lot of places. Headcount goes down even as output goes up.

• 2030+: new baseline. The skills that compound become table stakes for the role. The current top quartile becomes the median. A new top quartile emerges, and the cycle repeats. This is how every major tooling transition has gone — IDE/refactor tools, version control, cloud — and there's no particular reason this one is different in pattern, only in speed.

If I had to pick the single number that captures the situation: the comp premium for top-quartile AI-fluent engineers at major labs is now 50-70%, up from approximately zero in 2023. That number is the live ammunition in the career-planning question.

What I'm not saying

I'm not saying "everyone will be fine if they just use Cursor." I'm saying the people most likely to be on the right side of the distribution are the ones who treat this as a serious skill-development question right now, not in 2028, not "when things stabilize." Things aren't stabilizing — the slope of the gap is steepening.

I'm also not saying AI is "doing the work." It's still you doing the work. The work just looks different. It looks more like editing and reviewing than typing. It rewards taste more than throughput. It punishes vibes-driven approvals more than it ever did.

The 10x engineer is back. They're not 10x because they're working harder. They're 10x because they're the leverage point for a layer of automation that didn't exist three years ago, and they've gotten good at being that leverage point. That skill is learnable, and it's the highest-return thing you can develop this year.

Further reading

• GitHub Octoverse 2026 — PR velocity by percentile — the underlying productivity data
• Latent Space — Cursor internal metrics episode — the leak with the user-cohort spread numbers
• Anthropic — engineering levels in the agent era — what compounding looks like inside Anthropic
• Stratechery — the productivity power law — Ben Thompson on the market-structure implications
• Patrick McKenzie — career strategy when the curve stretches — the career-planning lens