AI Coding Agents: More Code, Same Software Output
AI coding agents are churning out code at a pace that would make any engineering manager's head spin. But a new study from Harvard researchers, using data from Jellyfish covering 300 million work events across 700+ firms, reveals a stubborn truth: more code doesn't mean more software. The efficiency gains from AI are being swallowed by a human review bottleneck, leaving firm-level output largely unchanged. Here's what that means for teams betting big on agentic coding.
The 30% Code Surge Meets a 49% Review Slowdown
When firms introduce AI coding agents, total lines of code jump 30%, commits rise 20%, and pull requests increase 23%. Sounds like a win—until you look downstream. The average time from pull request submission to merge balloons 49%. The share of pull requests requiring changes nearly doubles, and comments per pull request climb 35%. Human reviewers are drowning in AI-generated output, and the review process itself becomes the new bottleneck. The net effect? No statistically significant change in resolution rates for Jira-tracked issues and epics. Firms are producing more code, but not more finished software.
Humans Still Dominate Review, AI Can't Save Itself
You might think AI could help review its own code, but the study found that impact marginal so far. By March 2026, 80% of measured firms used some form of AI code review, yet AI agents were responsible for only 23.3% of review comments and 10.8% of pull requests. Humans remain the primary reviewers, and their workload has increased—the share of workers performing code reviews rose 14% after AI agents arrived. Meanwhile, the researchers found no significant employment changes attributable to AI, and no compositional shift in the size or complexity of tracked issues. The review bottleneck is real, and it's human.
Why This Matters for Your Dev Team
This isn't an argument against AI coding tools—it's a reality check on where the friction lives. If your team is adopting agents, expect the bottleneck to shift from writing code to reviewing it. That means investing in review capacity, better tooling, and clear guidelines for when to deploy agents. The study suggests these trade-offs could improve as teams gain experience, but for now, the double-edged sword is sharp: faster coding, slower merging. The practical takeaway? Measure your review cycle time, not just code generation speed, and staff accordingly. Otherwise, you're just generating more work for your senior engineers.
Key Takeaways
- AI coding agents increase code volume by 30% but don't improve software output at the firm level.
- Code review time balloons 49% after AI agents are introduced, becoming the new bottleneck.
- Humans still handle 76.7% of review comments and 89.2% of pull requests, despite AI review tools.
- No significant employment changes or issue resolution improvements were attributed to AI.
- Teams should focus on review capacity and cycle time, not just code generation speed.
Source: Ars Technica • 🇺🇸 San Francisco
Keep Reading
