← back
What Data from 20m Pull Requests Reveal About AI Transformation — Nick Arcolano, Jellyfish
Takeaway
Interactive coding tools are delivering measurable 2x throughput and faster cycles without quality regressions, while fully autonomous agents remain pre-production at most companies.
Summary
- Jellyfish analyzed ~20M PRs from ~1000 companies / 200K devs (June 2024–present); median company AI adoption rose from 22% to ~90% of coding time in a year.
- Companies generating >=50% of code with AI grew from ~2% last summer to ~50% today; autonomous agents (Devin/Codex) used by only 44% of companies and <2% of merged PRs — still mostly experimentation.
- Productivity: going from 0% to 100% AI adoption correlates with ~2x PR throughput and ~24% lower cycle time; PR size up 18% (mostly new lines), files-touched stable.
- Quality: no significant regression in code quality observed so far at the cohort level.
ai-productivitycode-generationdata
Original description
Engineering teams are spending millions on AI coding tools, but most have no idea what's actually working. Without hard data, you're flying blind – unable to tell which teams are actually using AI effectively. But what if you had access to workflow data from 200,000 engineers and 20 million pull requests across a thousand companies? In this talk, we'll share insights from usage data spanning the entire AI engineering ecosystem. We've observed significant productivity gains at scale, including a 2x increase in PR throughput and 24% faster cycle times on average. You'll learn what "good" adoption looks like (hint: autonomous agents aren't there yet), what productivity gains are possible, and what side effects to expect. More importantly, we'll explore why some teams don't see these gains. We'll show how your code architecture" is a critical, often overlooked factor. --- Socials: - LinkedIn: https://www.linkedin.com/in/arcolano/ - X (Twitter): https://x.com/arcolano - GitHub: https://github.com/arcolano - Company: Jellyfish (https://jellyfish.co)