Study Finds AI Coding Agents Increase Code Volume but Not Software Output Due to Review Bottlenecks

A recent study by Harvard University researchers Fiona Chen and James Stratton analyzes the impact of AI coding tools on software development efficiency across more than 700 firms. Using aggregated analytics from Jellyfish covering 300 million work events and 700,000 employees from 2021 through March 2026, the researchers found that while AI coding agents significantly increase raw code production, they do not lead to a proportional increase in completed software features.

Code Generation vs. Output

The introduction of AI coding agents led to measurable increases in code volume:

• 30% increase in total lines of code generated.
• 20% rise in the number of total commits.
• 23% increase in pull requests.

However, these metrics did not translate into improved software output. The resolution rate for Issues and Epics (wholesale software features) tracked by tools like Jira showed no statistically significant change after AI tools were introduced. Furthermore, the study found no compositional shift in the size or complexity of these issues.

The Code Review Bottleneck

The discrepancy between code generation and software output is attributed to downstream constraints in the production process, specifically the code review phase:

• Average review process time (from pull request submission to merge) increased by 49%.
• The share of pull requests with changes requested nearly doubled.
• The number of comments per pull request increased by 35%.

To manage this increased workload, the share of workers performing code reviews rose by 14%. Despite the adoption of AI for code review by 80% of measured firms, humans remain responsible for the vast majority of this work, with AI agents accounting for only 23.3% of review comments and 10.8% of pull requests.

Employment and Future Outlook

The study found no evidence that firms reduced employment or increased overall software output by using AI coding tools. The researchers could not attribute significant employment changes to AI when cross-referencing active worker data with LinkedIn profiles. While 95% of firms in the study have implemented AI coding agents, the authors note that many are still in a learning phase regarding optimal deployment. The current trade-off suggests that gains in coding speed are largely counteracted by increased human effort in reviewing and refining AI-generated code.