A multi-layered detection framework analyzing 180 million Git repositories reveals that single-signal methods significantly underestimate the prevalence of generative AI coding agents, missing up to 97% of activity. The study identifies over 320,000 commits per month from agents like Claude Code, which dominates silent adoption through configuration files rather than bot accounts.
- Multi-method detection identified 850,157 Claude Code commits, whereas bot-account lookup recovered only 3.3% of these, indicating a 30x relative-recall gap for single-signal estimates.
- Every detection pattern was hand-validated with 495 labels, using per-cell precision and Wilson confidence intervals to ensure accuracy across snapshots from December 2024 to April 2026.
- Claude Code leads in adoption with 886,122 commits across 17,295 projects, primarily through silent configuration-file-only adoption affecting 21,078 projects.
- An independent pull-request census (AIDev) captures nearly disjoint populations from commit detection, missing 79% of Claude Code adopters and essentially all Codex adopters.
- Cloud agents deployed via pull requests (Codex, Cursor) surface as feature work, while in-editor agents (Claude Code, OpenHands, Aider) surface as maintenance, showing that work profiles follow deployment modes rather than the tool itself.
The authors conclude that no single detection channel is representative of AI agent usage, emphasizing the need for multi-method approaches to accurately assess their impact on the open-source supply chain.