AI writes code faster than you can review it
Othman Shareef · June 20, 2026 · 6 min read
In GitHub’s controlled experiment, developers using Copilot completed a coding task 55% faster than the control group. That was 2022-era tooling: autocomplete, not agents. Nobody has published the companion number, because it doesn’t exist: reviewing a line of code is exactly as expensive as it was in 2019. Human reading didn’t get a model upgrade.
Amplifiers amplify both directions
There’s a line making the rounds in engineering leadership circles: AI is an amplifier: teams with strong practices get stronger, teams with weak ones get worse. Review is where that cashes out. A team that already wrote small PRs with good descriptions now ships even faster, because review keeps up. A team that tolerated 1,500-line mixed-concern PRs now produces them twice as often, and the review queue, not the roadmap, sets their velocity.
The failure modes are predictable. Diffs grow, because generating is cheap and splitting is work, the trend that made us write how big should a PR be? Reviewers rubber-stamp, because honest review of the incoming volume is impossible and something has to give. Or reviewers hold the line and become the constraint, absorbing the team’s frustration for doing their job. Pick your dysfunction.
“Just review less” isn’t the answer: understanding is the product
It’s tempting to conclude review is now overpriced ceremony: let CI and the AI reviewer bots gate merges. But code review was never only defect-hunting. Microsoft’s research on review at scale found its biggest delivered values include knowledge transfer and shared understanding, not just bugs caught. In the agent era that function gets more important: when a model writes the code, review is the moment a human takes ownership of it. Skip that, and you ship a system nobody on the team understands, a debt with compounding interest.
The three places to attack the asymmetry
- Shrink the unit of review. Size caps, stacked PRs, splitting norms: everything in the data says this is the highest-leverage move. Small changes are the one thing that makes every other fix cheaper.
- Move review upstream. The author (or their agent’s supervisor) reads the diff before it becomes a PR (pre-PR review), and an AI first pass clears mechanical noise. Human reviewers should see changes a human has already vouched for, with a checklist like our AI-code review checklist applied.
- Stop taxing the reviewer. If review is the scarce resource, every lost scroll position, collapsed file, and buried thread is waste of the exact thing you’re short on. This is the lever we work on at Pyor: making the irreducible human part (reading and understanding a diff) as fast as the surface allows, especially for the large PRs that keep coming.
The uncomfortable conclusion
Most teams’ effective shipping speed is now set by review throughput, not authoring speed, whether they’ve noticed or not. The teams that internalize that will redesign around it: smaller units, upstream review, surfaces that respect reviewer attention. The rest will keep buying faster code generation and wondering why the cycle time didn’t move. The bottleneck doesn’t care how fast you fill the queue.