AI agents are increasingly writing open source software code
Everyone talks about AI agents. But how much AI-written code actually ships to production?
To find out, we mapped code contributions (pull requests) with AI agents such as Claude Code or Codex as co-authors for critical open source software code. The result for these 1 million repositories is an increase from 2% in October 2025 to 10% in May 2026. In absolute numbers this is 6x growth in 6 months, as can be seen in the graph below between September and March.
The share of rejected compared to accepted code contributions has also decreased, and so the pull requests don’t seem to necessarily be lower-quality contributions. However, arising challenges in the field include malicious attacks on critical packages, and an increase of invalid bug reports. The higher productivity has an effect on maintaining software packages and reviewing all this new code: developers have to do more manual checks of proposed code than ever. Many important codebases are maintained by one or two individuals and the fatigue of these volunteers watching over critical open source software is far from ideal. In response to this, we're seeing companies like GitHub ship new features to help maintainers manage the higher volume of contributions they're seeing.
Which repositories are critical is determined here by the Open Source Security Foundation. The list includes the backbones of many software packages such as Node.js, npm and Raspberry Pi as well as software from Apple, Microsoft, amongst others. The increase in code contributions is not only visible in direct contributions from agents: human-authored work went up from around 300.000 to 400.000 in just a few months, with generative AI tools likely supporting many of these contributions. However, since March 2026 we observed a decline in human-authored code for 2 months in a row. Although seemingly unrealistic, at this pace agentic AI would write more pull requests than humans within a year.
And the critical open source projects aren’t the only ones that experience accelerated growth: In his keynote at Computex 2026, Jensen Huang highlighted GitHub data showing indicators like merged pull requests, code commits and the creation of new repositories are reaching unprecedented heights.
The interest in AI agents is not limited to writing code. The share of AI papers focusing on AI agents also increased rapidly last year, moving from about 1% to more than 4% in the first quarter of 2026, and it’s still rising as indicated in the figure below.
Whereas there is a dip in code production in April every year, we clearly expect the amount and the share of code contributions from AI agents only to go up. Even though reduced token subsidies appeared around April, the growth of agent share of writing code overall is remarkable.
The rise of agentic AI in coding isn’t isolated from the challenges the open source software field is facing. Pressure is concentrating around a small number of individuals in an underfunded field. While we celebrate the potential of collaborative software development, we should also make sure to safeguard it, especially now that interest is ramping up.