AI Coding Agents Scaling Volume, Not Software Delivery
AI-generated image · Bay Street Wire
A study of hundreds of firms reveals that gains in AI-generated code are being neutralized by a massive surge in human review requirements.
The industry's push toward AI coding agents is producing a surge in raw output that fails to translate into actual software delivery, according to reporting from Ars Technica.
Researchers Fiona Chen and James Stratton from Harvard University analyzed aggregated analytics from Jellyfish, covering 300 million work events across more than 700 software development firms and 700,000 employees between 2021 and March 2026. The study found that while AI agents increase total lines of code by 30%, total commits by 20%, and pull requests by 23%, these metrics do not improve the resolution rate of software features (Issues and Epics) tracked in tools like Jira.
Instead, the efficiency gains are absorbed by a human review bottleneck. Following the introduction of AI agents, the average time between a pull request submission and its merge increased by 49%. The researchers noted that the share of pull requests requiring changes nearly doubled, while comments per pull request rose by 35%.
To manage this increased cognitive load, the share of workers performing code reviews grew by 14%. Despite 80% of firms using some form of AI code review by March 2026, humans remained the primary reviewers; AI agents accounted for only 10.8% of pull requests and 23.3% of review comments. Consequently, the researchers found no significant evidence that these tools increased overall software output or reduced employment levels.