Developers now spend more time reading code than writing it.
Faros AI tracked 22,000 developers across 4,000 teams over two years of increasing AI adoption — telemetry data, not surveys. Their finding: 11.4 hours per week reviewing code. 9.8 hours writing it. The primary activity of a software developer in 2026 is reading output. Not producing it.
Both Blades
Two things happened simultaneously.
Output accelerated. Faros measured it across their full dataset: epics per developer up 66%, task throughput up 33.7%, PR volume up 16.2%. LinearB's analysis of 8.1 million pull requests across 4,800 organizations found AI-assisted PRs run 2.6 times larger than manual ones. AI coding tools did what they promised. Generation got fast.
Attention contracted. Gloria Mark's longitudinal research at UC Irvine tracked screen-focus duration across two decades: 2.5 minutes in 2004, 75 seconds by 2012, 47 seconds by 2016 — replicated five times, never recovered. The human side of the verification equation was already degrading before AI multiplied the volume.
Between these two blades sits the thing that determines whether software works: the capacity of an organization to verify what was generated. Here is what happened to it.
Bugs per developer up 54%. Incident-to-PR ratio up 242.7%. Code churn — code rewritten shortly after being written — up 861%. Pull requests merged without any review up 31.3%. Median review time up 441.5%. Commit-to-production lead time up 480.4%. All Faros telemetry.
LinearB found the bottleneck in sharper relief. Agentic pull requests — generated by autonomous coding agents — wait 5.3 times longer before a reviewer touches them. Acceptance rate for AI-generated code: 32.7%. For code written by humans: 84.4%. More generated. Most rejected.
CircleCI captured it from the delivery side: feature-branch throughput up 59% year over year. Main-branch throughput — code reaching production — down 7%. More written. Less shipped.
The Complication
Individual developers adapt.
METR's follow-up study found that the same developers who were 19% slower with AI tools in 2025 are now 18% faster in early 2026. The learning curve runs 30 to 100 hours of deliberate practice before net gains appear. METR calls this "only very weak evidence" due to selection effects — but the direction is consistent.
I've been too quick to dismiss this. My earlier framing — "structural, not transitional" — became a reflex, not a conclusion. Fifteen months of data isn't enough to close that question. Individual developers learning to use AI tools effectively is genuine and documented.
But individual adaptation doesn't rescue the system.
Where the Scissors Cut
The scissors cut organizations, not developers.
A developer who gets 18% faster generates more code that enters a review queue already 441% slower. The gain is personal. The cost is collective. And the collective cost compounds: LinearB found junior reviewers spend 15 minutes per PR and accept 31.9%. Senior reviewers spend 38 minutes and accept 23.7%. Experience sees more problems. The better the reviewer, the worse the bottleneck.
Code review is a human-to-human function — quality gate, knowledge transfer, architectural oversight. It cannot be parallelized the way generation can. One developer with an AI agent produces five pull requests before lunch. One senior engineer still reviews two.
Every AI coding tool reports generation metrics. Tokens produced. Completions accepted. Lines suggested. Time saved. These arrive automatically, and the numbers are always good.
No tool reports how long a human spent verifying what the AI produced. No dashboard shows the rejection rate. No leaderboard ranks organizations by verification throughput. The thing that determines whether software actually works is the thing nobody sells and nobody measures.
Generation scales with compute. Verification scales with humans. For individuals, the answer is learn the tools. For organizations, there is no answer yet — just a growing queue of plausible code that nobody has time to read.