Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Playbooks · Coding tasks

180M repos: most agent work is unattributed

Written November 8, 2026

A 180M-repo census found 850,157 Claude Code commits — bot lookups found 28,154. Why GitHub counts miss most AI coding, and what that means for reviewers.

What the census did

A paper by Arsham Khosravani and Audris Mockus, submitted in June 2026, ran a multi-method census across World of Code, more than 180 million Git repositories. It combined four detection methods: configuration-file scanning, commit-message analysis, author-identity matching, and bot-signature lookup, classifying agent traces into four behavioural types. Every detection pattern was hand-validated, with 495 labels checked and per-cell precision reported.

The headline comparison is between two ways of counting the same thing. Multi-method detection found 850,157 Claude Code commits in one snapshot. Bot-account lookup, the signal most adoption studies rely on, recovered only 28,154 of them: 3.3 percent, a 30-fold relative-recall gap. The paper's conclusion is that single-signal prevalence estimates are biased low by at least that factor. Across snapshots from December 2024 to April 2026, commit-attributed agents generated over 320,000 commits per month. Claude Code leads, with 886,122 commits across 17,295 projects, and dominates silent, configuration-file-only adoption, with 21,078 projects showing agent configuration and no other trace.

Why bot counts miss most of it

No single method captures more than a fraction of agent activity, and the methods capture different populations. Compared against an independent pull-request census, the two channels capture nearly disjoint agent populations: a pull-request census misses 79 percent of commit-detected Claude Code adopters and essentially all Codex adopters. The work profile follows deployment and detection mode rather than the tool itself. Pull-request-deployed cloud agents, such as Codex and Cursor, surface as feature work, while commit-deployed in-editor agents, such as Claude Code, OpenHands, and Aider, surface as maintenance. No single channel is representative, because each channel sees the work its deployment mode produces.

That is the mechanism behind the gap. A bot account exists only where the agent acts as a bot: a cloud agent opening a pull request, or a commit authored under a bot identity. An agent running inside an editor, committing as the developer, leaves no bot account behind. It may leave a configuration file, a commit message pattern, or nothing at all. Counting bot accounts therefore counts one deployment mode, and mistakes it for the whole population.

What a reviewer should do instead

Limits

Related playbooks

Sources

Back to: Coding playbooks hub · Also: Fingerprinting AI coding agents from 33,580 PRs · METR: Anthropic's AI R&D sped up about 1.9 times · BrokenML: a benchmark of broken machine-learning jobs