AI agent pull requests on GitHub, 2026
How many pull requests coding agents open on public GitHub, how many get merged, how fast, and how many are merged without anyone submitting a review. Counted from the GitHub search API for January to September 2026, for four agents we can identify reliably, with all public pull requests as a baseline.
Data collected 2026-10-24. Free to reuse under CC BY 4.0. Download the counts (CSV), the merge-time sample (CSV) and the spot checks (CSV).
What the data shows
- Claude Code's attribution line is everywhere. Pull requests carrying it were 1.9% of all public pull requests opened in January 2026 and 18.2% of those opened in September: 4,147,718 of 22,802,650.
- Most agent pull requests get merged, and fast. Of those opened January to June 2026, 91.9% of Claude Code's, 75.3% of Copilot's, 58.3% of Devin's and 55.0% of Jules's were merged, against 74.5% for all public pull requests. In our sample the median merged agent pull request was merged 10 min (Claude Code) to 35 min (Jules) after it was opened.
- Almost none had a formal review. Between 90.1% and 96.7% of merged agent pull requests had no review submitted on GitHub, against 85.1% of all merged public pull requests.
- Devin's pull requests stay open longest. 3560900.0% of Devin's pull requests opened January to June 2026 were still open when we counted, against 682620200.0% for all public pull requests. Of those that had been decided, 82.7% were merged.
- Volumes move a lot. Copilot coding agent pull requests peaked in March 2026 at 317,324 and were 60,047 in September. Pull requests from branches named
copilot/fell the same way, so this isn't a change of bot account. Jules opened 102,089 in January and 5,127 in February; its high volume ran through December 2025 and all of January (see the spot-check CSV). Search can't tell us why. Because January dominates Jules's first-half total, its pooled rates mostly describe January.
None of this says the code was good or bad. It says how often agent pull requests are accepted on public GitHub, how quickly, and how rarely anyone submits a formal review first. Read the limitations below before quoting a number.
Key numbers
Pull requests opened January to June 2026, so every one has had at least three months to be merged or closed.
What happened to the pull requests
| Opened January to June 2026 | Opened | Merged | Closed, not merged | Still open | Merged of opened | Merged of decided |
|---|---|---|---|---|---|---|
| GitHub Copilot coding agent | 1,216,928 | 916,847 | 166,586 | 133,495 | 75.3% | 84.6% |
| Devin | 120,944 | 70,561 | 14,774 | 35,609 | 58.3% | 82.7% |
| Jules | 138,038 | 75,903 | 30,744 | 31,391 | 55.0% | 71.2% |
| Claude Code | 4,745,416 | 4,361,147 | 282,607 | 101,662 | 91.9% | 93.9% |
| All public pull requests | 72,533,122 | 54,003,904 | 11,703,016 | 6,826,202 | 74.5% | 82.2% |
"Merged of decided" leaves out pull requests that are still open: it is merged ÷ (merged + closed without merging). "Still open" is opened minus merged minus closed. We checked that against GitHub's own open count for one agent and month (Copilot, March 2026), and it matched exactly.
Month by month
Later months have had less time for pull requests to be merged or closed, so their rates can still rise. The volume swings a lot from month to month; we report it as counted and haven't tried to explain each swing.
| Pull requests opened | 2026-01 | 2026-02 | 2026-03 | 2026-04 | 2026-05 | 2026-06 | 2026-07 | 2026-08 | 2026-09 | Total |
|---|---|---|---|---|---|---|---|---|---|---|
| GitHub Copilot coding agent | 193,255 | 246,269 | 317,324 | 240,515 | 151,959 | 67,606 | 68,325 | 62,423 | 60,047 | 1,407,723 |
| Devin | 5,284 | 7,462 | 11,711 | 27,310 | 36,193 | 32,984 | 26,651 | 33,111 | 39,912 | 220,618 |
| Jules | 102,089 | 5,127 | 6,959 | 5,999 | 8,451 | 9,413 | 13,730 | 10,469 | 10,124 | 172,361 |
| Claude Code | 165,393 | 321,684 | 608,077 | 749,616 | 1,165,683 | 1,734,963 | 2,318,042 | 2,086,518 | 4,147,718 | 13,297,694 |
| All public pull requests | 8,922,238 | 9,999,871 | 12,313,096 | 12,407,585 | 13,927,830 | 14,962,502 | 17,243,755 | 18,922,432 | 22,802,650 | 131,501,959 |
How fast they merge
| Merged PRs sampled | Sample size | 25th percentile | Median | 75th percentile | Merged within 1 hour | Merged within 24 hours |
|---|---|---|---|---|---|---|
| GitHub Copilot coding agent | 1,200 | 5 min | 16 min | 1.4 h | 71.3% | 93.2% |
| Devin | 1,059 | 3 min | 14 min | 1.6 h | 71.0% | 92.1% |
| Jules | 1,079 | 1 min | 35 min | 8.1 h | 55.6% | 87.0% |
| Claude Code | 1,200 | 2 min | 10 min | 1.2 h | 73.8% | 93.1% |
| All public pull requests | 1,200 | 0 min | 4 min | 2.0 h | 71.7% | 88.8% |
Time to merge is from opening to merge, for merged pull requests only. It comes from a sample, not every pull request: up to 100 merged pull requests per agent on each of 12 days (the 1st and 15th of January to June 2026), 5,738 in all. GitHub returns at most 100 results per request, newest first, so each day's sample leans to pull requests opened late that day (UTC).
Merged without a review
The share of merged pull requests with no review submitted on GitHub (search qualifier review:none). Comments don't count as a review, and neither does a person reading the diff before pressing merge. So this measures formal review, not whether anyone looked.
| Opened January to June 2026 | Merged | Merged, no review submitted | Share |
|---|---|---|---|
| GitHub Copilot coding agent | 916,847 | 886,856 | 96.7% |
| Devin | 70,561 | 63,544 | 90.1% |
| Jules | 75,903 | 68,776 | 90.6% |
| Claude Code | 4,361,147 | 4,138,618 | 94.9% |
| All public pull requests | 54,003,904 | 45,961,975 | 85.1% |
What we haven't measured yet
- Reverts within 30 days. Search can't tell us which pull requests were later reverted. Measuring it means reading the commits after each merge, repository by repository, which the public API's rate limits don't allow at this scale. Not measured.
- CI pass rate. Check results are a separate request per pull request, and many small repositories have no CI at all. Not measured.
- Pull requests with and without tests. Knowing whether a pull request changed test files needs its file list, one request each. A sample big enough to mean anything per agent would take days at the unauthenticated rate limit. Not measured.
We will add these if we can measure them honestly, and say so here when we do.
Method
Which pull requests count as an agent's
We only counted an agent where its own documentation gives a way to recognise its pull requests: a bot account that opens them, or a fixed line the agent adds to the description by default. Each method finds a floor, not a total: it misses pull requests where the person changed the default.
| Agent | Search qualifier | How it identifies the agent | What it misses |
|---|---|---|---|
| GitHub Copilot coding agent | author:app/copilot-swe-agent | Pull requests opened by the Copilot coding agent's GitHub app, which shows as the author "Copilot". GitHub Docs: Managing agent sessions | Work from the agent that a person copies into a pull request of their own. |
| Devin | author:app/devin-ai-integration | Pull requests opened by the Devin.ai Integration GitHub app, the account Devin uses when it opens pull requests as Devin. Devin Docs: GitHub integration | Pull requests from workspaces set to open them as the user rather than as Devin. |
| Jules | author:app/google-labs-jules | Pull requests opened by the Jules GitHub app. Jules's docs say that when you publish a PR from Jules, Jules appears as its creator. Jules Docs: Reviewing code | Pull requests a person opens by hand from a branch Jules published. |
| Claude Code | "Generated with Claude Code" in:body | Pull requests whose description contains the line Claude Code adds by default: "🤖 Generated with [Claude Code](https://claude.com/claude-code)". The author is the person's own account. Claude Code Docs: attribution.pr setting | Pull requests where the attribution line is turned off or edited out, and ones a person writes the description for. |
| All public pull requests | is:public | Every pull request in a public repository: the baseline the agents are compared with. It includes the agents' own pull requests and those of dependency bots. GitHub Docs: Searching issues and pull requests | Nothing within public repositories. Private repositories are not searchable without access. |
The exact queries
For each agent and each month of 2026 from January to September, by the date the pull request was opened (UTC), four searches of https://api.github.com/search/issues, read for their total_count. For Copilot in January:
is:pr author:app/copilot-swe-agent created:2026-01-01..2026-01-31 is:pr author:app/copilot-swe-agent is:merged created:2026-01-01..2026-01-31 is:pr author:app/copilot-swe-agent is:closed is:unmerged created:2026-01-01..2026-01-31 is:pr author:app/copilot-swe-agent is:merged review:none created:2026-01-01..2026-01-31
The merge-time sample, for one day, sorted by creation date, newest first, 100 per page:
is:pr author:app/copilot-swe-agent is:merged created:2026-01-01
Every query we ran is in the counts CSV, next to its result: 180 queries. The months pooled for the headline rates are 2026-01 to 2026-06. The collection script is packages/console/app/script/research/ai-agent-prs.ts in our repository; it needs nothing but Bun and a network connection, and takes about 45 minutes without a GitHub token.
Limitations
- Public repositories only. Private repositories, where most company code lives, are not searchable. The results describe public GitHub, which includes many personal projects, experiments and test repositories.
- Floors, not totals. Each method misses pull requests where the person turned off or edited the default. Claude Code's count includes any description that contains the line, whoever wrote it.
- A merge isn't a verdict on quality. In a personal repository the person who asked the agent usually merges its pull request. A high merge rate can mean the work was good, or that nobody else was asked.
- The baseline includes the agents. "All public pull requests" also contains the agents' own pull requests and those of dependency bots, which merge often.
- Search counts are GitHub's. We report the totals GitHub's search returns. GitHub flags a search that didn't finish; none of ours were flagged. Counts can change slightly as repositories are made private or deleted.
- No people or repositories. We publish counts and anonymous merge times only. The study names tools, never developers or projects.
Use the charts
Each chart is a standalone image you can embed, with a link back here:
What happened to pull requests opened January to June 2026
<a href="https://aicodeops.dev/research/ai-agent-pull-requests-2026"><img src="https://aicodeops.dev/research/charts/ai-agent-prs-2026-outcomes.svg" alt="What happened to pull requests opened January to June 2026" width="760" loading="lazy" style="max-width:100%;height:auto"></a> <p><a href="https://aicodeops.dev/research/ai-agent-pull-requests-2026">AI agent pull requests on GitHub, 2026</a> by <a href="https://aicodeops.dev/">AI CodeOps</a>, CC BY 4.0</p>
Share of pull requests merged, by month opened (2026)
<a href="https://aicodeops.dev/research/ai-agent-pull-requests-2026"><img src="https://aicodeops.dev/research/charts/ai-agent-prs-2026-merge-rate-by-month.svg" alt="Share of pull requests merged, by month opened (2026)" width="760" loading="lazy" style="max-width:100%;height:auto"></a> <p><a href="https://aicodeops.dev/research/ai-agent-pull-requests-2026">AI agent pull requests on GitHub, 2026</a> by <a href="https://aicodeops.dev/">AI CodeOps</a>, CC BY 4.0</p>
Median time from opening to merge (sampled merged PRs)
<a href="https://aicodeops.dev/research/ai-agent-pull-requests-2026"><img src="https://aicodeops.dev/research/charts/ai-agent-prs-2026-time-to-merge.svg" alt="Median time from opening to merge (sampled merged PRs)" width="760" loading="lazy" style="max-width:100%;height:auto"></a> <p><a href="https://aicodeops.dev/research/ai-agent-pull-requests-2026">AI agent pull requests on GitHub, 2026</a> by <a href="https://aicodeops.dev/">AI CodeOps</a>, CC BY 4.0</p>
Merged pull requests with no review submitted (opened January to June 2026)
<a href="https://aicodeops.dev/research/ai-agent-pull-requests-2026"><img src="https://aicodeops.dev/research/charts/ai-agent-prs-2026-merged-without-review.svg" alt="Merged pull requests with no review submitted (opened January to June 2026)" width="760" loading="lazy" style="max-width:100%;height:auto"></a> <p><a href="https://aicodeops.dev/research/ai-agent-pull-requests-2026">AI agent pull requests on GitHub, 2026</a> by <a href="https://aicodeops.dev/">AI CodeOps</a>, CC BY 4.0</p>
Why this matters for teams adopting coding agents, and what to put between an agent's pull request and your main branch: AI code review, quality gates and how AI CodeOps reviews every change before a person approves it.
Frequently asked questions
Which coding agents does this cover?
GitHub Copilot coding agent, Devin, Jules and Claude Code: the agents whose pull requests we could identify on public GitHub using a method each vendor documents. All public pull requests are included as a baseline.
Why are OpenAI Codex and Cursor missing?
In the pull requests we looked at, both were opened from people's own accounts, and we found no identification method in either vendor's documentation that we could rely on, such as a bot author or a fixed attribution line. Branch-name prefixes exist, but anyone can use them, so we left both out rather than guess.
Does a high merge rate mean the code was good?
No. Most of these pull requests are in small public repositories where the person who asked the agent also merges the result, often without a formal review. A merge says someone accepted the change, not that it was correct.
Can I use the data?
Yes. The CSV files, the charts and the numbers on this page are licensed CC BY 4.0. Use them anywhere, with a link back to this page. The collection script is published too, so you can re-run it.
Cite this
AI CodeOps (2026). AI agent pull requests on GitHub, 2026. https://aicodeops.dev/research/ai-agent-pull-requests-2026
The data and charts on this page are licensed CC BY 4.0: reuse them anywhere, with a link back to this page.