Have We Solved AI Coding's Biggest Bottleneck: Code Review?
Over the past two years, AI has largely solved code generation. But code review capacity hasn't scaled at the same pace. The bottleneck has shifted from writing code to manually reviewing it.
A recent study by Carnegie Mellon and Stanford, analyzing 196,212 pull requests (PRs) at one company, found that per-developer PR creation increased to 2.09× the baseline, while the review workload per reviewer nearly doubled.
In other words, AI has made developers significantly more productive, but it has also created a growing queue of PRs that still depend on human reviewers. Several industry reports are pointing to the same conclusion.
Are AI Code Review Agents the missing piece?
We may now be seeing the next wave of AI engineering tools: AI Code Review Agents.
One of the most interesting examples comes from PostHog (Product Analytics and Engineering Company), which recently published an engineering article titled "Stop Being the Code Review Bottleneck."
Their internal AI code review agent, StampHog, has already given the final approval for roughly one-third of the PRs independently handling approximately 1,600 PRs.
For perhaps the first time, we're seeing an engineering organization allowing an AI agent to approve a meaningful share PRs.
What makes their approach interesting?
Rather than replacing human reviewers, the goal is to automate routine reviews and allow engineers to focus on other high-value engineering work.
Some notable engineering practices adopted by PostHog include:
AI reviews routine PRs and escalates only higher-risk changes.
The agent that generates the code never reviews its own work.
Large changes are broken into smaller, incremental PRs, making automated validation more reliable.
Requirements and implementation intent are validated before code generation begins, reducing downstream review effort.
The real insight isn't that AI will replace code reviewers. It's that AI review and human review are complementary - not competitive.
This isn't an isolated experiment. GitHub recently announced that Copilot Code Review surpassed 60 million code reviews by March 2026 - a 10× increase since April 2025.
Of course, AI code review isn't free. Running AI code review agents introduces additional inference and token costs, so organizations will need to balance those costs against the engineering productivity they unlock.
I am curious to hear how other engineering teams are approaching this.
References
Posthog Article: https://newsletter.posthog.com/p/code-review-tips
Carnegie Mellon Stanford Study: https://arxiv.org/pdf/2607.01904v1
GitHub Code Review Growth: https://github.blog/ai-and-ml/github-copilot/60-million-copilot-code-reviews-and-counting
