Jellyfish across 100 million PRs: agents touch over a quarter of pull requests, but full software factories barely register and people still approve almost every merge
Jellyfish's October report, covered by R&D World, finds agents opened or committed to 27.7% of PRs at the median company in September, while merges approved on AI review alone sat at 0.5% (8.5% at the 90th percentile) and end-to-end factories opened 9.4% of PRs only at the 99th-percentile company. Engineers who kept editing agent context files merged more.
Brian Buntz (R&D World) reports on Jellyfish's October AI Engineering Trends report, built from activity data across more than 1,300 companies, 276,000 developers and 100 million pull requests. In September, AI agents opened or committed code to 27.7% of PRs at the median company, up from about 2% in January. The median share of code written by AI tools was 29%. Claude Code spend reached a median of $63 per active developer per week, about $252 a month, and $1,666 a month at the 90th percentile.
The review numbers move much more slowly. At the median company, 0.5% of merged PRs went through on AI review with no human input. At the 90th percentile that was 8.5%, up from about 4% in January. Jellyfish defines a software factory as a pipeline where agents pick, write, review and ship the work with little human involvement. By that definition, factories opened 0.0% of PRs (after rounding) even at the 90th-percentile company, and 9.4% at the 99th, up from about 1.7% in March.
Jellyfish's head of research, Nicholas Arcolano, frames the ceiling as human attention: if a person babysits every step, "we've hit the limit," and the way through is to hand off whole tasks, starting with simple, safe ones. One correlate stands out. Engineers who edited agent context files six or more times merged 2.3 more PRs a week than those who never did, controlling for AI usage, and each doubling of context-file work was associated with 29% more merged PRs. In a survey of 1,417 developers, 88% said AI tools helped their work, but only 62% said their organization manages the risks of AI-generated code well. Jellyfish sells an engineering-management platform, and these are observational figures from its customers.
Why it matters for code review
Talk about overnight factories is loud, but this is a baseline from 100 million PRs: agents write a large share of changes, and a person still approves almost every merge. The gap between those two curves is where review queues pile up, as Assembled saw before its auto-reviewer and as Yu and colleagues measured in reviewers who approve more and comment less. The context-file result suggests the steady, unglamorous work of maintaining the harness is what moves merges, more than any single session.