AI Coding Agents Add 30% More Code and Zero Additional Resolved Issues, Study Finds — Bottleneck Is Review, Not Generation
A working paper by Fiona Chen and James Stratton, titled “Artificial Intelligence in the Firm: Bottlenecks in Software Production,” presents the first large-scale empirical study of AI coding agent deployments at enterprise scale. The findings, covered by Ars Technica on October 9, 2026, show a consistent pattern: agents dramatically increase code production metrics while leaving output metrics unchanged.
The Numbers
| Metric | Change After Agent Adoption |
|---|---|
| Lines of code generated | +30% |
| Total commits | +20% |
| Pull requests opened | +23% |
| Code review time | +49% |
| Revision requests | Nearly doubled |
| Resolved issues | No change |
The paper’s current version is dated August 4, 2026. The research covers enterprise-scale deployments; the exact firm and dataset specifics are in the working paper.
Where the Bottleneck Is
The study’s core finding is that software production is not bottlenecked at the code generation stage. Introducing AI coding agents removes friction from writing code but does not remove friction from the review, quality-assurance, and merge process that must follow.
Review times increased 49% in firms that adopted agents. The authors attribute this to a combination of increased volume hitting review queues and a per-PR increase in the thoroughness required — AI-generated code introduces failure modes that reviewers have not built intuition for over years of reading human-written code. Revision requests nearly doubling is consistent with reviewers catching more issues per PR rather than reviewers becoming more demanding.
The result is a shift in the production function: code generation becomes cheap and fast, while code review becomes the scarce resource. Firms that treated agent adoption as a force-multiplier on engineering headcount without accounting for review capacity found the gains absorbed by review overhead.
What This Means for AI Review Tools
The study’s implicit challenge is directed at the second wave of AI coding tools — automated code reviewers — rather than at coding agents themselves. If review is the constraint and human review time has increased 49%, the economic argument for AI-assisted code review is now significantly stronger than it was before agent adoption.
Several tools — including GitHub Copilot’s review features, Cursor’s background analysis, and purpose-built products from Qodo and CodeRabbit — are positioned to address this. The Chen-Stratton data provides the empirical grounding for that investment. None of those tools were independently evaluated in this study.
Context: Prior Research
The findings differ from a June 2026 MIT study covering 100,000 developers across three tool generations, which found AI coding agents raised commits by 180% and releases by 30%. The contrast is notable: the MIT study tracked individual developer productivity metrics; Chen and Stratton track firm-level output. It is possible both are correct — individual developers write more code and open more PRs, while the firm-level constraint (review capacity) absorbs the gain before it converts to shipped software.
The policy implication for engineering leaders: adopting coding agents without expanding review capacity, either through headcount or tooling, may produce a backlog rather than a throughput gain.