---
title: 'In the News: October 12, 2026 (Extra 3)'
description: 'A Harvard working paper finds agent adoption raised code output but also review time, with no significant change in resolved Jira work.'
canonical_url: 'https://darkfactory.dev/news/2026-10-12-extra-3'
markdown_url: 'https://darkfactory.dev/news/2026-10-12-extra-3.md'
collection: news
date_published: '2026-10-12T23:10:00-04:00'
date_modified: '2026-10-12T23:10:00-04:00'
---

# In the News: October 12, 2026 (Extra 3)


One item this cycle. A Harvard working paper on firm-level data finds that adopting coding agents raised output measured in code and pull requests, and also raised the time the output spends in review.

## 1. Harvard working paper: agents add code and PRs, but also review time, and no measurable gain in resolved Jira work

**[Artificial Intelligence in the Firm: Bottlenecks in Software Production](https://fion.ac/jellyfish.pdf)** · Fiona Chen and James Stratton, Harvard University · Working paper, current version August 4, 2026

The authors use proprietary Jellyfish data on 718 firms and 725,938 workers from January 2021 to March 2026. Per the paper, adopting AI agents is associated with about 30% more lines of code, 20% more commits and 23% more pull requests per worker-month. Review time rises by 3.45 days (s.e. 1.60) against a 7.03-day baseline, about 49%. The authors write that "the share of PRs with changes requested increases by 0.12 (s.e. = 0.01) relative to a baseline mean of 0.13", and report 35% more review comments per PR. The paper finds no statistically significant effect on Jira issues or epics resolved per worker-month, and its abstract says "productivity gains do not fully pass through to changes in software output or employment." Ars Technica [covered the study](https://arstechnica.com/ai/2026/10/ai-coding-agents-generate-more-code-but-not-more-software/) and it reached Hacker News on October 12 (314 points, 275 comments at the time of the front-page reading).

The estimates compare firms by adoption of agents, not individual developers, and adoption timing is not random: larger firms adopted earlier, so the authors caution against reading the numbers causally. Agent adoption is inferred from bot accounts and commit signatures. The data is proprietary, the sample is limited to Jellyfish clients who opted in, and the paper is unrefereed.

**Why it matters:** If your agents raise PR volume, plan for the review queue to absorb part of the gain. Measure review time and change-request rate next to throughput, since this paper finds those moved while resolved tickets did not. Treat it as one observational dataset, not a settled result.