AI Assisted Analysis / Code review
In the third quarter of 2026, openclaw merged about 2,700 pull requests a week. 88% of them merged without a human comment from anyone other than their author.
The pull request is still how code gets into these projects. But the page around it, where people used to question a change before it landed, has gone quiet. In the projects that build with agents most heavily, it has become a delivery vehicle: changes pass through it, and nobody stops to talk.
That is the far end of a pattern we found in 26 popular open-source repositories. Projects that started in the AI era merge far more pull requests and talk about each one far less. Established projects haven't sped up much, but they are moving the same way: fewer human comments per pull request, bigger pull requests, and more of the conversation written by bots.
We count comments, not attention. Silence on GitHub doesn't prove nobody read the change. It does mean the page no longer carries the review.
We counted every merged pull request from January 2024 to September 2026, 240,653 in total before leaving out dependency and release automation, and read the comments and reviews on a random sample of 143,815 of them (143,297 after leaving out codex's weeks from mid-May 2026, as the CSV does). The data is downloadable as CSV.
What we measured #
We compared two groups of repositories:
- Established (16): react, next.js, vscode, cpython, django, rust, kubernetes, svelte, vite, astro, tailwindcss, deno, bun, mermaid, playwright and xyflow (React Flow). All were large before AI coding tools were common.
- AI-native (10): opencode, pi, codex, gemini-cli, openclaw, crush, qwen-code, kilocode, cline and goose. All build AI coding tools; cline and goose started in 2024, the rest in 2025.
For every repository we tracked two numbers per quarter:
- Merges per week: merged pull requests, counting those opened by people and by agent accounts, and leaving out dependency updates, release automation and backports.
- Human comments per PR: conversation comments, inline review comments and review summaries written by people other than the PR's author. Bot commands such as
/lgtmor@bors r+and templated messages (the same opening posted 20 or more times by one person, such as landing receipts) don't count. Bot comments are counted separately and split in two:AI review (CodeRabbit, Copilot, Gemini Code Assist, Claude, Codex, Greptile and similar) andautomation (CI status, test, coverage and size reports, preview deployments, changeset and CLA checks).
Our baseline is the first quarter of 2025. That isn't a world without AI: Copilot and Cursor were in wide use, and Claude Code launched that February. But the established repositories looked the same throughout 2024. Their median human comments per PR stayed between 1.8 and 2.1 in every quarter, so the first quarter of 2025 is a fair "before", and the charts start there.
The map #
Established projects discuss each PR less; the busiest AI-native ones barely discuss them.
Merges per week against human comments per merged PR. Established projects move from Q1 2025 to Q3 2026; AI-native projects are shown in Q3 2026.
Read it from two directions.
Down: 12 of the 16 established repositories now get fewer human comments per pull request than in early 2025. React stayed flat; Kubernetes, Rust and next.js went up.
Right: 12 of the 16 merge more per week than they did, but for most the change is small. The median established repository went from 30 to 32 merges a week. The biggest moves are vscode (195 to 363 a week) and bun (56 to 180); mermaid and xyflow doubled from a small base.
The AI-native repositories sit further right and lower. opencode merged 303 pull requests a week in the third quarter, with 0.08 human comments per PR. openclaw merged about 2,700, with 0.18. They are not all alike, though. qwen-code merged 268 a week with 1.3 human comments per PR, and 56% of its pull requests were approved by someone other than the author; it also gets 3.9 AI review comments per PR, most of them posted from maintainers' own accounts. Being AI-native doesn't force a project to skip review. Most of the busiest ones leave little of it on GitHub.
Established teams: the conversation thinned #
Across the 16 established repositories, the typical repository changed like this between the first quarter of 2025 and the third quarter of 2026:
Pull requests got about 90% bigger and drew about 40% fewer human comments. The change starts in the last quarter of 2025, the quarter Anthropic released Claude Opus 4.5 together with Claude Code updates; the quarters before barely moved. The timing lines up, but it doesn't prove cause: other agents and models shipped in the same months.
Bots now write two of every three comments on established projects' pull requests.
Comments per merged PR by who wrote them, average of the 16 established repositories. In Q1 2025 bots wrote a little over one in three.
Count bot comments along with people's and review looks busier than in early 2025: the average established repository went from 4.3 comments per merged PR to 6.0. About half of the added bot comments are automation (CI status, reports, preview deployments, checks), which grew from 1.6 comments per PR to 2.7 on average; in the median repository AI review is close to zero (0.03 per PR), so automation is nearly all of its bot traffic. The other half is AI review, and it is concentrated: more than one AI review comment per 20 pull requests in only 6 of the 16 repositories in the third quarter of 2026. Where it's used it is busy: bun gets 9.5 AI review comments per pull request, tailwindcss 4.1, mermaid 3.0 and vscode 2.6.
Three established repositories saw the biggest surges in volume, and each traded discussion for speed in its own way:
- bun went from 56 merges a week to 180, and from 1.28 human comments per PR to 0.66; 80% of its pull requests now merge with none. (Bun joined Anthropic in December 2025.) In the third quarter of 2026, 79% of its merged pull requests came from
robobun, an account the Bun team uses for automated pull requests. - deno spiked to 110 merges a week in the second quarter of 2026, then returned to 40. Most of the spike came from accounts we classified as agents, such as
divybotandnathanwhitbot, which opened about half of that quarter's pull requests; across all its pull requests that quarter, 69% merged with no human comment. - vscode nearly doubled to 363 merges a week. It already had few human comments per PR (0.43, now 0.34), and 85% of its pull requests still merge with none.
Three repositories kept the most human review. Django still gets 5.3 human comments per PR, and only 14% of its pull requests merge with none. Rust gets 4.6, with 24% at none. Kubernetes gets 9.6 even with its /lgtm and /approve commands set aside, and only 11% merge without a human comment. All three require a named human sign-off to merge.
AI-native teams: volume without conversation #
In the third quarter of 2026, the median AI-native repository merged 76 pull requests a week, against 32 for the median established one: more than twice as many. The average per repository is even more lopsided, 392 against 78, because openclaw alone merged about 2,700 a week.
More pull requests merge without a single human comment.
Share of merged PRs with no comment from anyone but the author, each repository weighted by how many PRs it merged.
Weighted by volume, 83% of AI-native pull requests merged without a human comment in the third quarter, up from 50% a year earlier. In the median AI-native repository the figure is 58%, close to the established median of 54%. The difference is where the volume goes: the two busiest AI-native projects are ones where almost nobody comments.
Pull requests are getting bigger, fastest in AI-native projects.
Lines changed (added plus deleted) in the median merged PR, median repository per quarter.
AI review is where the AI-native projects differ most. In the third quarter of 2026, 8 of the 9 measured that quarter (codex excluded, see below) had an AI reviewer commenting on pull requests, and the median repository got 2.8 AI review comments per merged PR against 0.9 from people. Automation is light there: 0.1 comments per PR.
AI-native pull requests are also getting bigger quickly. The median one changed 158 lines in the third quarter, nearly twice the established median of 85. And an approval from someone other than the author is less common: 53% of merged pull requests in the median AI-native repository, against 65% for established ones.
One AI-native repository needed special handling. Merges on GitHub in openai/codex nearly stopped in mid-May 2026, and from late August almost every pull request comes from copyberry, an account that syncs changes from an internal repository (its branches are named codex-internal-to-codex-oss). The review happens there, out of view, so we plotted codex at its first-quarter 2026 numbers and left the weeks from mid-May out.
How this fits other research #
Faros AI's AI Productivity Paradox report (July 2025, 1,255 teams and over 10,000 developers) found that developers on teams with high AI adoption "merge 98% more pull requests", while "PR review time increases 91%" and average PR size rises 154%. Our established repositories show the size effect, about 90% bigger, but not the throughput: their median merge rate barely moved. We didn't measure review time, but in the median repository the median time from opening a pull request to merging it rose from about 14 hours to 26 while human comments fell.
A study of 33,228 pull requests in vLLM and SGLang (Li, Ye and Olowoniyi, August 2026) reports that "PR comment density increased 4.2x in vLLM and 3.8x in SGLang, with bot comments contributing an estimated 15-20% of the increase." That doesn't contradict what we found. Count every comment, bots included, and review looks busier than ever; that is the full height of the bars in Figure 2. Count only people other than the author, and the conversation is shrinking in most of the repositories we measured. In established repositories about half of the added bot traffic isn't review at all, but CI and report automation.
What it means for your team #
As pull requests got bigger and, in some projects, far more numerous, human comments per pull request fell in most repositories we measured. That comment is also the thing that told the next maintainer why a change exists.
- Bot findings are not shared understanding. An AI reviewer can tell you what might be wrong with a change. It doesn't leave a person who knows what the change does.
- The PR is often the only record. When a pull request is bigger and nobody discusses it, its description is all the next reader gets. Make that record good, and cheap to produce.
- Decide where a human must sign off. Django, Rust and Kubernetes kept the most human review, and all three require a named human sign-off to merge. Pick the paths in your code that deserve that, and let the rest go faster.
- Track your own two numbers. Merges per week and human comments per PR are easy to pull from the GitHub API. If the first rises while the second falls, ask where the review is happening now.
aictrl.dev · Explain Change
The pull request page, reinvented. #
When nobody has time to comment, the PR has to explain itself. Explain Change turns every pull request into interactive diagrams: the whole change at a glance, and one click from any step to the exact code behind it. High-level understanding and line-level depth, in the same view.
Example: aictrl-dev/cli#129, which makes the CLI enforce a JSON Schema on the model's final answer. The flowchart follows aictrl's explanation of that pull request; the sequence and state views are illustrations. Every code excerpt is the pull request's own code.
Limits #
- This shows change, not its cause. The groups are defined by project, not by measured AI use. All the AI-native projects build coding tools, and some established changes have other causes, like release cycles or team changes. Only a few percent of pull request descriptions say an AI tool wrote them, so we couldn't measure AI authorship directly.
- Review outside GitHub isn't counted. Discord threads, pair work and internal review leave no comments. Codex shows how completely review can move out of view.
- AI review versus automation is decided by account. Every bot comment on a sampled pull request is classed by its author: AI reviewers (CodeRabbit, Copilot, Gemini Code Assist, Claude, Codex, Greptile and similar) or automation. AI-written CI triage, such as Playwright's flake verdicts, counts as automation. AI review posted from a person's own account looks human to us. We looked for it everywhere: openings structured like AI review output, mentions of AI tools, bulk inline reviews, and by reading samples of comments. We found it in one project, qwen-code, where four maintainers post signed AI review from their own accounts; we count those as AI review. We found no sign of it in the projects that kept the most human review, such as Kubernetes, Rust and next.js, but AI-assisted text without markers can't be ruled out.
- Bots are detected by account type and name. We classified every bot-like author with five or more merged pull requests since 2025 as automation (left out) or agent (kept), checking unclear ones against their pull requests. Some automated accounts are registered as ordinary users; we found the large ones, such as
robobun, by hand. - Sampling. Weekly merge counts are exact. For comments, size and approvals we listed every merged pull request in each week and drew a random sample of 100 (every one, in smaller weeks), 143,815 in total and 143,297 after leaving out codex's weeks from mid-May 2026, weighted by each week's merge count. Weeks run Monday to Sunday, so the third quarter of 2026 ends on 27 September.
- Approvals are GitHub reviews. "Approved" means a GitHub review in the approved state. Projects that approve through bot commands, such as Rust's bors or Kubernetes' Prow, score low on it.
All 26 repositories: Q1 2025 and Q3 2026 #
| Repository | Group | Merges/week Q1 2025 | Merges/week Q3 2026 | Human comments/PR Q1 2025 | Human comments/PR Q3 2026 | No human comment, Q3 2026 |
|---|---|---|---|---|---|---|
| denoland/deno | Established | 31.5 | 39.5 | 1.77 | 0.96 | 60% |
| django/django | Established | 17.0 | 13.2 | 6.59 | 5.25 | 14% |
| kubernetes/kubernetes | Established | 63.4 | 78.7 | 8.48 | 9.63 | 11% |
| mermaid-js/mermaid | Established | 3.5 | 7.5 | 2.33 | 1.15 | 60% |
| microsoft/playwright | Established | 39.7 | 44.5 | 1.35 | 1.07 | 61% |
| microsoft/vscode | Established | 194.7 | 363.4 | 0.43 | 0.34 | 85% |
| oven-sh/bun | Established | 56.2 | 180.3 | 1.28 | 0.66 | 80% |
| python/cpython | Established | 116.9 | 142.4 | 4.59 | 2.44 | 49% |
| [react/react](https://github.com/react/react) | Established | 29.2 | 13.5 | 1.41 | 1.39 | 44% |
| [rust-lang/rust](https://github.com/rust-lang/rust) | Established | 169.3 | 197.2 | 4.12 | 4.59 | 24% |
| sveltejs/svelte | Established | 16.5 | 10.3 | 1.69 | 0.99 | 34% | | tailwindlabs/tailwindcss | Established | 16.4 | 4.1 | 2.26 | 0.79 | 49% | | vercel/next.js | Established | 96.6 | 100.9 | 0.98 | 1.57 | 58% | | vitejs/vite | Established | 14.6 | 22.0 | 2.11 | 1.12 | 60% | | withastro/astro | Established | 20.2 | 24.1 | 2.98 | 1.39 | 51% | | xyflow/xyflow | Established | 3.9 | 9.0 | 0.65 | 0.26 | 81% |
| [aaif-goose/goose](https://github.com/aaif-goose/goose) | AI-native | 60.5 | 54.0 | 1.65 | 0.92 | 49% |
| [anomalyco/opencode](https://github.com/anomalyco/opencode) | AI-native | — | 303.2 | — | 0.08 | 97% |
| [charmbracelet/crush](https://github.com/charmbracelet/crush) | AI-native | — | 16.8 | — | 1.58 | 51% |
| [cline/cline](https://github.com/cline/cline) | AI-native | 22.6 | 76.2 | 1.40 | 1.02 | 61% |
| [earendil-works/pi](https://github.com/earendil-works/pi) | AI-native | — | 28.5 | — | 0.72 | 55% |
| [google-gemini/gemini-cli](https://github.com/google-gemini/gemini-cli) | AI-native | — | 10.1 | — | 2.04 | 58% |
| [Kilo-Org/kilocode](https://github.com/Kilo-Org/kilocode) | AI-native | — | 103.1 | — | 0.67 | 75% |
| [openai/codex](https://github.com/openai/codex) | AI-native | — | 177.9 (Q1 2026) | — | 1.48 | 62% |
| [openclaw/openclaw](https://github.com/openclaw/openclaw) | AI-native | — | 2665.2 | — | 0.18 | 88% |
| [QwenLM/qwen-code](https://github.com/QwenLM/qwen-code) | AI-native | — | 268.2 | — | 1.30 | 45% |
— means the repository had no full quarter of merged pull requests then. Human comments exclude the PR author and bots.
Where does your team sit on Figure 1? If you've measured your own two numbers, we'd like to hear what you found.
Sources #
- Pull request data: GitHub GraphQL API, 26 repositories, merged pull requests from 1 January 2024 to 27 September 2026, collected by aictrl.dev on 15 October 2026. Per-repository quarterly data (CSV) .
- Faros AI, The AI Productivity Paradox Report 2025 , 23 July 2025.
- Jiada Li, Xuesong Ye, Olamide Olowoniyi, Engineering Signals of Human-AI Collaboration in the Agentic Coding Era: A Longitudinal Analysis of 33,228 Pull Requests from vLLM and SGLang , arXiv:2608.13884, 14 August 2026.
4. Bun, [Bun is joining Anthropic](https://bun.com/blog/bun-joins-anthropic) , 2 December 2025.
5. Anthropic, [Claude 3.7 Sonnet and Claude Code](https://www.anthropic.com/news/claude-3-7-sonnet) , 24 February 2025.
6. Anthropic, [Introducing Claude Opus 4.5](https://www.anthropic.com/news/claude-opus-4-5) , 24 November 2025.