On this page
On 1 October 2026, openclaw landed 699 commits on its main branch. Hermes Agent landed 642. For comparison, requests has about 6,500 commits in its entire 15-year history.
These are two of the most-starred projects on GitHub. Both were built in the last 15 months, and both are coding-agent projects. I wanted to see what development at this speed looks like from the inside, first in the git history and then in the code.
Commit pace
Commits per day on each project's default branch, from GitHub, for the week before I wrote this:
| Date (UTC) | openclaw | Hermes Agent | deepseek-harness |
|---|---|---|---|
| 28 Sep | 592 | 671 | 165 |
| 29 Sep | 482 | 271 | 151 |
| 30 Sep | 591 | 202 | 98 |
| 1 Oct | 699 | 642 | 19 |
| 2 Oct | 625 | 504 | 40 |
| 3 Oct | 571 | 595 | 8 |
| 4 Oct | 380 | 290 | 0 |
| 5 Oct | 426 | 484 | 0 |
Scroll the table sideways to see every column.
openclaw's whole history is about 106,000 commits since its first commit in November 2025. Hermes Agent has nearly 50,000 since July 2025. deepseek-harness has about 20,700 since June 2026, which averages out at around 175 a day.
Authors
A day of 600 to 700 commits sounds like the work of a big team, and the author counts show a mix of a wide contributor base and a small, very fast core. On 1 October, Hermes Agent's 642 commits came from 81 different authors, and its two most active authors wrote about half of them. openclaw's 699 came from 72 authors, and a single maintainer account wrote 446 of them, about 64%.
The git history alone can't tell you how much of this code a model wrote. Of those two days' commits, 8 of Hermes Agent's and 6 of openclaw's carry an AI co-author trailer ("Co-Authored-By: Claude" and similar), which is about 1%. I expect the real share is higher, because the trailer is optional. A study that measures "AI code" from commit metadata ends up measuring which authors leave the trailer in.
Size
| openclaw | Hermes Agent | deepseek-harness | |
|---|---|---|---|
| Lines of code, code files only (at our last index) | 4.2 million | 1.8 million | 0.9 million |
| Files | 13,293 | 8,758 | 12,578 |
| Indexed on | 26 June | 19 August | 28 September |
Scroll the table sideways to see every column.
openclaw reached 4.2 million lines of code about seven months after its first commit. For a sense of scale, that's larger than Grafana (2.2 million lines of code, 13 years of history) and much larger than React (0.6 million).
Two of these indexes are old by these repos' standards: openclaw has added tens of thousands of commits since we last indexed it. Treat the sizes as "at least".
Code shape
Long functions get split
At our index, Hermes Agent's agent loop was a single 6,591-line function and openclaw's run function was 4,972 lines. Both have since been taken apart into small functions, which are 48 and 35 lines today. I'd guess the pace is why: when every new feature has to pass through one function, a fast team runs into that limit early. I wrote about this in the longest functions post.
Documentation follows the language
We count the share of functions and classes that have a docstring or doc comment. Hermes Agent (Python) is at 36%, against a 24% median for popular Python repos. deepseek-harness (TypeScript) is at 21% and openclaw (TypeScript) at 2%, against a 7% median for popular TypeScript repos. I think types do much of the documenting in TypeScript, which would explain the low median.
Published research
These numbers fit what the published research would predict.
- GitClear analysed 211 million changed lines from 2020 to 2024 and found that "moved" code, a sign of refactoring, fell from about 25% of changed lines to under 10%, while copy-pasted code rose, so more new code was written and less old code was reshaped.
- METR ran a randomized trial in early 2025 where experienced open-source developers took 19% longer with AI tools on their own repos. Their February 2026 follow-up with newer tools pointed to a speedup instead, but METR says plainly that the new data is an unreliable signal, because developers who rely on AI declined to take part or avoided tasks where it helps most.
- DORA's 2025 report found that AI adoption now goes with higher delivery throughput but still with lower delivery stability.
- CodeRabbit reviewed 470 open-source pull requests and found about 1.7× more issues in AI-co-authored PRs than in human-only ones.
And one result in the other direction, from our own data: when we traced 112,000 commits back to the bugs they caused, commits made with coding agents were not more bug-inducing than the rest.
My own read, which these numbers don't prove, is that at 500 commits a day the hard part moves from writing code to the work around it: review, testing, knowing which file does what, and knowing why it was written that way. repowise is built for that work. It keeps a map of the codebase (what depends on what, who owns which files, which decisions shaped them) and serves it to coding agents over MCP, so the agent writing commit 501 knows what the first 500 did. You can connect it to Claude or ChatGPT for any public repo.
What these numbers can't tell you
- Commit counts measure activity, so they say little about quality. Many of these commits are small, and some projects commit every step an agent takes. 600 commits from one team and 60 from another may be the same amount of work.
- The AI co-author trailer count is only a lower bound on the AI share.
- Our indexes of openclaw and Hermes Agent are from June and August. Both codebases have changed a lot since.
- The three repos are different products in different languages, indexed at different times. I've put them side by side to show scale, and they aren't ranked.
How we measured
Daily commit counts come from the GitHub API (/repos/<repo>/commits?since=…&until=…, committer date, UTC, default branch) on 6 October 2026. Authors and co-author trailers come from the same API for 1 October. Lines of code (code files only), file counts and documentation share come from each repo's latest repowise index. "Popular repos" means the 197 public repos in our index with 500+ stars and a health score.