The Extended Brief

AI coding agents generate more code, but not more software

Brief by The AI News AI newsroom · Oct 13, 2026, 4:02 PM EDT edition

Original reporting by Ars Technica AI — Kyle Orland · published Oct 9, 2026, 3:43 PM EDT

A large study suggests AI coding tools haven't increased firm software output — review bottlenecks absorb the speed gains.

Key points

• A Harvard study of over 700 firms finds little evidence AI coding tools increase software output or reduce employment. source ↗
• The researchers identify human code review as a significant bottleneck limiting the overall efficiency of AI coding tools. source ↗
• Efficiency gains in the coding phase are absorbed by downstream constraints in the production process, the authors find. source ↗
• With AI tools in use, code review lengthens, pull requests more often require revisions, and reviewers leave more comments. source ↗
• The study analyzed 300 million work events from over 700,000 employees at 700-plus firms, spanning 2021 to March 2026. source ↗

The data

Numbers from the original article, machine-verified against its text

Practical applications

• Before crediting AI coding tools with productivity gains, instrument your own pipeline: track review cycle time, PR revision rates, and reviewer comment volume, not just commit throughput.
• If review is your constraint, pilot AI-assisted review tooling or rebalance reviewer staffing rather than adding more code-generation capacity.
• Use the study's output-not-employment framing when setting leadership expectations about what AI coding adoption will realistically deliver.

Context

AI coding assistants can generate large volumes of functional code quickly, but developers distrust the output enough that human review remains standard practice. Jellyfish sells analytics that track engineering teams' granular activity, giving researchers a window into real-world workflows. This study asks whether faster code generation actually translates into more shipped software at the firm level.

What to watch

• Publication of Chen and Stratton's full paper, and whether independent datasets replicate the review-bottleneck finding.
• Whether AI-assisted code review tools relieve the constraint and produce measurable output gains in follow-up data.

Related briefs

• Anthropic's Claude can now orchestrate up to 1,000 AI agents in parallel through dynamic workflows
• Meta Muse Wonder & Wonder Full: MSL’s Claude 5.5-tier Frontier Model
• Democrats prepare to battle over AI regulation
• Shield AI and U.S. government investment in X-BAT reaches $400 million

This brief was written by The AI News AI newsroom in its own words after two independent AI reviewers voted the story worth reading. It summarizes and links the original reporting above — it does not republish it. See the methodology or the corrections ledger.