Code Grew 30%, But Shipping Didn't──Harvard Measured the "Traffic Jam That Moved to Review" Across 700 Companies
Harvard analyzed roughly 300 million development events across 700 companies. After adopting AI agents, code grew 30% but task resolution stayed flat. The bottleneck had shifted from "writing" to "reviewing and verifying."
Deploy a coding agent, and the amount of code it writes does indeed go up. Yet the thing users actually want──the speed of shipping working software──has barely changed. That's the number Harvard researchers Fiona Chen and James Stratton put on the table in a study released on October 9, 2026. It's a large-scale analysis covering more than 700 software companies and roughly 300 million development events (commits, pull requests, and the like) from 2021 through March 2026.
What went up was "how much gets written"; what stayed flat was "how much gets done"
After adopting agents, each company's development activity rose on average as shown below. Meanwhile, the resolution rate for Jira-tracked issues and larger features reportedly showed no statistically significant improvement.
| Metric | Change after adoption (average) |
|---|---|
| Lines of code | +30% |
| Number of commits | +20% |
| Number of pull requests | +23% |
| Issue/feature resolution rate | No significant improvement |
In other words, the first half of the production line (the code-writing stage) sped up, yet the volume shipped as finished product didn't grow. The researchers trace the source of this gap to the second half of the process.
The bottleneck moved from "writing" to "verifying"
The more code an agent churns out, the more the burden of reading it, testing it, and checking whether it fits the context piles up on the human side. The study found the weight of the review stage increasing as follows.
- The average time from when a pull request is submitted to when it's merged rose +49%
- The number of comments per pull request rose +35%
- The share of pull requests sent back for changes roughly doubled
What's easy to overlook is that much of this verification work still falls to humans. As of March 2026, about 80% of the companies measured were using some form of AI review, yet review comments generated by AI agents accounted for only 23.3% of the total, and pull requests opened by AI made up just 10.8%. The writing side can be handed to AI, but the verifying side still rests in human hands──that's the picture.
Why "it should be faster" doesn't show up in billing or delivery dates
From a management perspective, this result puts a check on the expectation that "adopting AI means instant cuts in development cost and faster releases." Even when code generation speeds up, the slack created upstream gets absorbed by downstream review and rework. The research team frames this as a state where, absent automated testing, verification mechanisms, and a reworking of the workflow, "only the volume of code grows, with no effect on release frequency or cost."
That said, it's a mistake to take this one-sidedly. The study also notes that there is scant clear evidence of companies cutting engineering jobs. This isn't a story of "AI taking jobs"; the reading closest to reality is that "the value-creating stage has moved from writing to review and verification, and that's where things are still clogged." Rather than denying productivity gains outright, it's best read as a warning that capturing the benefit requires rebuilding the downstream stages.
What teams can do on the ground to clear the jam
So what does it take to connect all that extra code through to shipping? Related industry data and operational know-how from various companies point to moves that lower the review burden earlier in the process.
Review with intent attached (review memory)
Attach to the pull request, in structured form, what changed and why, under what assumptions and risks, and how it was verified. This lets reviewers skip "reconstructing the background" and focus on the judgment itself.
Put the rules in the repository
Use a conventions file such as AGENTS.md or .cursorrules to spell out in advance which patterns to use, which shared utilities exist, and which constraints to honor. This reduces variance in the agent's output and lowers review friction.
Narrow the blast radius with staged releases
Use feature flags and gradual rollout to verify in small increments before exposing a change to everyone. Instead of demanding perfection in a single review pass, you can spread out the pressure of verification by limiting the blast radius and gathering facts as you expand.
All of these are about affecting "the burden humans bear in verifying" rather than "how much AI writes." When you measure the impact of adopting an agent, too, you risk misreading the real picture unless you look all the way to downstream metrics such as submit-to-merge time and rework rate──not lines of code or commit counts.
References: Harvard Study Finds AI Coding Agents Lift Code 30% but Fail to Raise Software Output (Zetik) / AI coding agents boosted code output without clear task-completion gains, study finds (Superpower Daily) / AI Coding Agents Generate More Code, Human Review Absorbs Gains (HyperAI) / The AI Coding Productivity Paradox 2026 (FeatBit)
