The era of hand-crafted code in the most popular AI tooling repositories is rapidly fading. A new data analysis of 52 of GitHub's most-starred AI-tooling projects, covering 324,088 commits, reveals that the median repository now has 16.9% of its recent activity signed by AI agents. This figure is a lower bound, as many agents do not leave digital fingerprints, but the trend is undeniable: 21 of the 52 repositories have at least 25% of their last 180 days of commits marked as AI-generated.

The Claude Code Monopoly

When looking at which tools are driving this shift, Anthropic’s Claude Code is the undisputed heavyweight. Of the 36,892 signed commits identified in the study, 85.3% were attributed to Claude. This dominance is partly due to Claude Code adding a trailer by default, while competitors like Cursor (855 commits), Copilot (797 commits), and Devin (325 commits) sign less frequently or use different methods. The data shows that 44 of the 52 repositories also commit agent instruction files like CLAUDE.md, indicating a deep integration of agentic workflows into the development lifecycle.

Quality Signals Are Confounded by Recency

The study attempted to determine if AI-heavy development leads to more bugs or higher churn by comparing these repos to 12 famous pre-Copilot projects. At the 180-day mark, AI-heavy repos showed a 34.1% fix commit rate compared to 17.3% in the baseline, and a lower code survival rate. However, the researchers caution that these gaps are likely driven by recency rather than AI usage itself. AI adoption is heavily concentrated in projects started in 2025 or later (correlation ρ = 0.72 with start date), and when comparing only recent projects, the difference in churn and survival between AI-signed and non-AI-signed commits vanishes.

No Link Between AI Usage and Popularity

Contrary to the hype that AI-driven development yields faster, more popular tools, the data shows no detectable link between the percentage of AI-signed commits and a repository's star count. The Spearman correlation is effectively zero (ρ = -0.05, p = 0.72). The ten most-starred repositories in the dataset range wildly from 0% to 73.6% AI-signed. This suggests that while agents are ubiquitous in the trenches of AI tooling development, they are not currently the primary driver of viral adoption or community trust in these projects.

Key Takeaways

  • Claude Code Dominance: Anthropic's tool accounts for over 85% of all AI-signed commits in the top 52 AI repos.
  • Recency Bias: High fix rates and churn in AI-heavy repos are likely artifacts of newer projects, not the AI agents themselves.
  • Popularity Disconnect: There is no statistical correlation between the volume of AI-written code and GitHub stars.
  • Self-Publishing Culture: 24 of the 52 repos publish their own benchmark harnesses, with several openly admitting to unfavorable results.

The Bottom Line

We are no longer debating if AI writes code in top-tier open source; we are debating if it changes the outcome. The data suggests the 'AI quality penalty' is a myth of newness, but the reliance on Claude Code is a single point of failure the community should watch.