Adoption is broad. Many contributors are committing agent-authored work, which is what a real organisational rollout looks like.
Assisted or autonomous
Assisted — a developer’s identity authored the commit and the agent added itself as co-author. The agent ran git commit during a working session, with a person there. Autonomous — a bot identity is the author, no human in the loop. Across the whole index, 93.7% of agent commits are assisted.
Agent share by month
Peak month: 17.5%. Bars are scaled to that peak, not to 100%.
Which agents
| Cursor Agent | Anysphere | 23 | 6.3% |
| GitHub Copilot | GitHub | 13 | 3.5% |
| Claude Code | Anthropic | 11 | 3% |
Verify it yourself
Sample commits that matched. Every one links to GitHub — check the trailers.
71d316d6Deprecate unused DeepSpeed features (#8490)Cursor Agent725811f9Add pin_empty helper for empty scratch destinations (#8332)Cursor Agenteebe24a6Route ZeRO/SuperOffload pin sites through accelerator pin_memory (#8256)Cursor Agent7075430aAdd DCO sign-off to release commits (#8527)GitHub Copilotbd6e9681Fix comms logger KeyError when log_name is omitted (#8267)GitHub Copilot29d0abbcAvoid collective token preparation for AutoEP DeepEP (#8423)GitHub Copilot431f8aceGive Muon's momentum the dtype of the gradient it is combined with (#8483)Claude Code1b892a75Read rope_theta from rope_parameters across Inference V2 (#8345)Claude Codeb5e000c4[muon] Keep the momentum out of steps the loss scaler discards (#8435)Claude Code
What was excluded
| Commits in window | 367 |
| Merge commits removed | −0 |
| CI / dependency bots removed | −0 |
| Considered | 367 |
Bots are excluded from the denominator entirely. Counting dependabot as “AI-written code” is the fastest way to produce a meaningless number.
How this was measured
Measured over February–August 2026, cloned directly. Detection is deterministic string matching against signatures agents write about themselves — no heuristics, no style analysis. Merge commits and CI bots are excluded from the denominator. The figure is a floor: squash merges strip trailers, so real usage is higher. Contributor counts are aggregate — no individual is named.
Run it yourself: npx qmmit-cli deepspeedai/DeepSpeed