AI Coding Agents Add Code, Not Software Output, Study Finds

A Harvard study covering more than 700 software firms finds that AI coding agents raise the volume of code produced by roughly 30 percent, yet deliver no statistically significant increase in finished software. The efficiency gained at the keyboard, the authors conclude, is being absorbed by a review process that humans still have to carry.

What the researchers measured

Fiona Chen and James Stratton of Harvard Economics worked with engineering-analytics firm Jellyfish according to Ars Technica, drawing on aggregated data covering more than 300 million work events from over 700,000 employees at 700-plus software companies between 2021 and March 2026. By combining measured AI usage with GitHub activity signals, they estimated when each firm introduced AI coding assistants, then AI coding agents, and compared output before and after.

The raw numbers look impressive. Adopting coding agents lifts total lines of code by 30 percent, commits by 20 percent, and pull requests by 23 percent on average. But the resolution rate of Jira-tracked issues and epics, meaning whole chunks of shipped functionality, did not move in a statistically meaningful way.

The review bottleneck

The gap shows up at code review. After agents arrive, the average time between a pull request being opened and merged balloons by 49 percent. The share of pull requests with changes requested nearly doubles, and comments per pull request rise 35 percent. Firms respond by pushing 14 percent more of their staff into reviewing work.

AI review tools are not closing that gap yet. By March 2026, 80 percent of the firms studied used some form of AI code review, but AI agents produced just 23.3 percent of review comments and a mere 10.8 percent of all pull requests. Humans still own the gate.

Why it matters for builders

For teams wiring agents into real workflows, this is a capacity story, not a model story. An agent that writes code 30 percent faster only pays off if the surrounding pipeline, review, CI, ownership and acceptance criteria, can absorb the extra volume. Automation that ends at code generation simply moves the queue downstream.

The practical takeaway is to instrument the whole loop. Track review latency and rework rate alongside generated volume, and treat the agent as one node in the pipeline rather than the finish line.

What to watch

The dataset cuts off in March 2026, and 95 percent of the firms now deploy coding agents, so the trade-off is still moving. Expect follow-up work to focus on whether review automation and tighter specs can reopen the bottleneck.

Full paper: Artificial Intelligence in the Firm: Bottlenecks in Software Production, with background from Jellyfish.