coderot

How many of Copilot’s agent PRs get merged? 51 C# repos, 12 months

In 20 of the most popular C# repos, pull requests made with AI were closed without being merged more often than their maintainers’ own in 13, and less often in 2. The clearest case is one agent, GitHub Copilot’s coding agent. The rest of the AI PRs were level with the maintainers’ own in 11 of 19 repos, and closed more often in 6.

Copilot’s coding agent had a third of its PRs closed. In the 11 repos where we could compare, 32% of the PRs it opened were closed unmerged, against 11% of the maintainers’ own: higher in all 11, from dotnet/efcore (19% against 7%) to dotnet/maui (52% against 21%).

It isn’t who uses the agent. We compared each person’s Copilot PRs with PRs they wrote themselves, in the same repo: for 54 of 73 people the Copilot PRs were closed more often, for 1 less. It isn’t PR size, and it didn’t get better over the year.

Agents people drive themselves are closer to their own PRs. PRs that maintainers made with Claude Code were closed as often as their other PRs in 6 of 7 repos, and those made with the Copilot CLI a little more often. The difference is between handing a task to an agent and working with one.

Everything coderot measures, and both our earlier reports, looks at merged code. This one looks at what wasn’t merged. All data is as of 18 and 19 November 2026.

All PRs
51 C# repos, 55,034 PRs opened from November 2025 to October 2026, merged, closed or still open
AI PRs
20 repos with 20 or more AI PRs and 20 or more maintainer PRs decided, 14,450 AI PRs
Copilot’s agent
11 repos with 20 or more of its PRs and 20 or more maintainer PRs decided, 3,424 Copilot PRs
Same people
73 people with 5 or more Copilot PRs they asked for and 5 or more of their own, in 12 repos

The short version

13 of 20repos closed AI PRs without merging more often than maintainers’ own PRs; 2 less often, 5 level

Clearest for one agent. Without Copilot’s coding agent, the other AI PRs were level in 11 of 19 repos and higher in 6. All AI PRs pooled: 17% closed against 11%, or 22% against 11% without SubtitleEdit, whose 3,219 AI PRs were closed 2% of the time.

32%of the PRs Copilot’s coding agent opened were closed without merging, against 11% of maintainers’ own PRs, in the 11 repos with enough of both to compare

Higher in every one of them. dotnet/runtime 27% against 14%, microsoft/aspire 39% against 11%, dotnet/roslyn 34% against 10%. Leaving out the 19% that were closed with no code in them, it’s still 27% against 11%.

54 of 73people had the Copilot PRs they asked for closed more often than PRs they wrote themselves, in the same repo; 1 had them closed less

It’s not who uses it. The median person’s Copilot PRs were closed 31% of the time, their own 12%. Counting only PRs of 1 to 400 lines, 42 people still had more closed and 9 fewer.

6 of 7repos closed PRs maintainers made with Claude Code as often as their other PRs; 1 less often. With the Copilot CLI, 5 of 10 more often

An agent you drive is closer to you. Pooled, Claude Code’s PRs were closed 3% of the time against 9%, but SubtitleEdit has 3,218 of its 4,133 PRs; without it, 6% against 9%. The Copilot CLI’s gap was smaller than the coding agent’s in every repo that had both.

8.4comments and reviews by people on a merged Copilot agent PR, on average, against 4.4 on a maintainer’s own merged PR

The ones that merge take more talk. That counts the requester steering the agent, and the author answering on their own PR. The 1,079 closed ones took 3,720 more.

How we measured

All AI PRs, then each agent

Each row compares a kind of AI PR with the maintainers’ own PRs in the same repos, counting repos with 20 or more decided PRs on each side.

PRs closed without merging Repos Higher Lower Level Pooled Pooled, without SubtitleEdit
All AI PRs 20 13 2 5 17% vs 11% 22% vs 11%
AI PRs but Copilot’s agent 19 6 2 11 13% vs 11% 17% vs 11%
Copilot’s coding agent 11 11 0 0 32% vs 11% same
The Copilot CLI 10 5 2 3 19% vs 12% same
Claude Code 7 0 1 6 3% vs 9% 6% vs 9%

Higher or lower means by a fifth or more and 3 points or more. Pooled adds up the PRs of every repo in the row, against the maintainers’ own in the same repos, so a big repo weighs more. SubtitleEdit weighs most: 3,219 AI PRs, nearly all Claude Code’s, closed 2% of the time, against only 41 maintainer PRs. So the last column leaves it out; it isn’t in the Copilot rows. Codex and opencode had enough in one repo each (PEASS-ng: 0% of Codex’s 20 closed; ILSpy: 4% of opencode’s 50, against 2%), and Cursor, Devin and the others in none, so they have no row of their own; they’re in the first two.

AI PRs were closed more often than maintainers’ own in most of these repos. Copilot’s coding agent was, in every one; the other AI PRs were level in most repos, but not all.

Repo by repo

Copilot’s coding agent opened PRs in 33 of the 51 repos, and most of them used it a little. 18 never had a PR from it (jellyfin, osu, Sonarr, Radarr and others), and 18 had 1 to 17, too few for a share. Four more had 20 to 55, but they’re run by one person who works through the agent or pushes directly, so there are almost no maintainer PRs to compare with (more on them under the table). That leaves 11, with 3,424 of the agent’s 3,718 PRs.

PRs closed without merging, per repodotnet/runtime: 27% against 14%; microsoft/aspire: 39% against 11%; dotnet/efcore: 19% against 7%; dotnet/maui: 52% against 21%; dotnet/roslyn: 34% against 10%; tui-cs/Terminal.Gui: 26% against 13%; dotnet/aspnetcore: 26% against 11%; microsoft/garnet: 30% against 15%; PowerShell/PowerShell: 52% against 8%; Flow-Launcher/Flow.Launcher: 92% against 14%; abpframework/abp: 35% against 4%PRs closed without merging, per repoCopilot’s coding agent against the same repo’s maintainers, PRs opened in 12 monthsCopilot’s coding agentMaintainers’ own PRs0%25%50%75%100%dotnet/runtimedotnet/runtime: maintainers’ own PRs, 14% of 2457 closed without mergingdotnet/runtime: Copilot’s coding agent, 27% of 1504 closed without merging27%microsoft/aspiremicrosoft/aspire: maintainers’ own PRs, 11% of 895 closed without mergingmicrosoft/aspire: Copilot’s coding agent, 39% of 672 closed without merging39%dotnet/efcoredotnet/efcore: maintainers’ own PRs, 7% of 230 closed without mergingdotnet/efcore: Copilot’s coding agent, 19% of 275 closed without merging19%dotnet/mauidotnet/maui: maintainers’ own PRs, 21% of 307 closed without mergingdotnet/maui: Copilot’s coding agent, 52% of 233 closed without merging52%dotnet/roslyndotnet/roslyn: maintainers’ own PRs, 10% of 1070 closed without mergingdotnet/roslyn: Copilot’s coding agent, 34% of 228 closed without merging34%tui-cs/Terminal.Guitui-cs/Terminal.Gui: maintainers’ own PRs, 13% of 207 closed without mergingtui-cs/Terminal.Gui: Copilot’s coding agent, 26% of 204 closed without merging26%dotnet/aspnetcoredotnet/aspnetcore: maintainers’ own PRs, 11% of 604 closed without mergingdotnet/aspnetcore: Copilot’s coding agent, 26% of 201 closed without merging26%microsoft/garnetmicrosoft/garnet: maintainers’ own PRs, 15% of 61 closed without mergingmicrosoft/garnet: Copilot’s coding agent, 30% of 30 closed without merging30%PowerShell/PowerShellPowerShell/PowerShell: maintainers’ own PRs, 8% of 581 closed without mergingPowerShell/PowerShell: Copilot’s coding agent, 52% of 29 closed without merging52%Flow-Launcher/Flow.LauncherFlow-Launcher/Flow.Launcher: maintainers’ own PRs, 14% of 59 closed without mergingFlow-Launcher/Flow.Launcher: Copilot’s coding agent, 92% of 25 closed without merging92%abpframework/abpabpframework/abp: maintainers’ own PRs, 4% of 645 closed without mergingabpframework/abp: Copilot’s coding agent, 35% of 23 closed without merging35%
The 11 C# repos with 20 or more decided PRs from Copilot’s coding agent and 20 or more from maintainers without AI, PRs opened November 2025 to October 2026. Closed without merging, as a share of merged plus closed. Together: 32% of 3,424 Copilot PRs against 11% of 7,116 maintainer PRs. Hover a dot for its counts.
Repo Copilot PRs decided Copilot closed Maintainers’ own closed Maintainer PRs decided
dotnet/runtime 1,504 27% 14% 2,457
microsoft/aspire 672 39% 11% 895
dotnet/efcore 275 19% 7% 230
dotnet/maui 233 52% 21% 307
dotnet/roslyn 228 34% 10% 1,070
tui-cs/Terminal.Gui 204 26% 13% 207
dotnet/aspnetcore 201 26% 11% 604
microsoft/garnet 30 30% 15% 61
PowerShell/PowerShell 29 52% 8% 581
Flow-Launcher/Flow.Launcher 25 92% 14% 59
abpframework/abp 23 35% 4% 645
All 11 3,424 32% 11% 7,116

In every repo that used Copilot’s coding agent enough to compare, its PRs were closed without merging more often than the maintainers’ own, two to three times as often in most.

Eight of the 11 are Microsoft’s, and dotnet/runtime alone has 1,504 of the 3,424 PRs. Without it, the other 10 pooled are at 35% against 10%.

Where one person runs the repo, it’s different. Four repos used the agent a lot but have too few maintainer PRs to compare, because their maintainer works mostly through it or pushes directly. There most of its PRs were merged: 43 of 43 in Tyrrrz’s DiscordChatExporter and YoutubeDownloader, 52 of 55 in mRemoteNG, and 15 of 24 in QuickLook. The gap above is in repos where a team decides what goes in.

The same people

A repo’s Copilot PRs could come from the people whose own PRs get closed more anyway. So for each person who asked the agent for 5 or more PRs, we compared those with the PRs they wrote themselves in the same repo.

Compared People Copilot PRs closed more Fewer Within 5 points
All their PRs 73 54 1 18
Only PRs of 1 to 400 lines, on both sides 67 42 9 16

The median person’s Copilot PRs were closed 31% of the time and their own 12%. Pooled over the 73, it’s 32% of 2,627 against 11% of 5,035.

Not size, and not getting better

Copilot’s agent PRs are about as big as maintainers’ own (a median of 73 lines merged, against 58), and the gap is there at every size:

PR size, lines added and deleted Copilot closed Maintainers’ own closed
50 or fewer 34% of 1,514 10% of 3,372
51 to 400 26% of 1,345 11% of 2,317
Over 400 36% of 565 15% of 1,427

By the month a PR was opened, Copilot’s share closed was between 26% and 39% every month, with no trend after March. The maintainers’ own stayed between 7% and 14%.

PRs closed without merging, by the month they were opened2025-11: Copilot 37%, maintainers 10%; 2025-12: Copilot 33%, maintainers 12%; 2026-01: Copilot 34%, maintainers 12%; 2026-02: Copilot 39%, maintainers 14%; 2026-03: Copilot 27%, maintainers 14%; 2026-04: Copilot 28%, maintainers 12%; 2026-05: Copilot 26%, maintainers 12%; 2026-06: Copilot 30%, maintainers 10%; 2026-07: Copilot 29%, maintainers 8%; 2026-08: Copilot 27%, maintainers 8%; 2026-09: Copilot 27%, maintainers 7%; 2026-10: Copilot 27%, maintainers 11%PRs closed without merging, by the month they were openedThe 11 repos of the chart above, all their PRs togetherCopilot’s coding agentMaintainers’ own PRs0%25%50%Nov 25Jan 26MarMayJulSep2025-11, Copilot’s coding agent: 37% of 426 closed2025-12, Copilot’s coding agent: 33% of 337 closed2026-01, Copilot’s coding agent: 34% of 427 closed2026-02, Copilot’s coding agent: 39% of 434 closed2026-03, Copilot’s coding agent: 27% of 373 closed2026-04, Copilot’s coding agent: 28% of 305 closed2026-05, Copilot’s coding agent: 26% of 251 closed2026-06, Copilot’s coding agent: 30% of 217 closed2026-07, Copilot’s coding agent: 29% of 235 closed2026-08, Copilot’s coding agent: 27% of 133 closed2026-09, Copilot’s coding agent: 27% of 160 closed2026-10, Copilot’s coding agent: 27% of 126 closed27%2025-11, maintainers’ own PRs: 10% of 665 closed2025-12, maintainers’ own PRs: 12% of 588 closed2026-01, maintainers’ own PRs: 12% of 806 closed2026-02, maintainers’ own PRs: 14% of 801 closed2026-03, maintainers’ own PRs: 14% of 749 closed2026-04, maintainers’ own PRs: 12% of 692 closed2026-05, maintainers’ own PRs: 12% of 622 closed2026-06, maintainers’ own PRs: 10% of 576 closed2026-07, maintainers’ own PRs: 8% of 457 closed2026-08, maintainers’ own PRs: 8% of 449 closed2026-09, maintainers’ own PRs: 7% of 408 closed2026-10, maintainers’ own PRs: 11% of 303 closed11%
Each point is the PRs opened that month, merged or closed by 18 and 19 November 2026, when they were read. Copilot’s share stayed between 26% and 39%, with no trend after March.

What the closed ones look like

Of the 1,079 Copilot PRs closed in the 11 repos, against the 798 maintainer PRs closed:

Closed PRs Copilot’s agent Maintainers’ own
No code in them (0 lines) 19% 2%
Closed within a day, with at most 1 comment 24% 26%
Closed after 3 or more comments and reviews 27% 30%
Closed by a stale-PR bot 8% 5%
Other 21% 36%
Still a draft when closed 85% 47%
Median days open 1 4.4

We read 40 of the Copilot ones. They’re of five kinds:

Most closed Copilot PRs weren’t argued over. They were tried, looked at, and dropped. About a quarter took real review before they were.

What it costs reviewers

People’s comments and reviews per PR Copilot’s agent Maintainers’ own
Merged PRs, average 8.4 4.4
Merged PRs, median 5 2
Closed PRs, average 3.4 3.4
Closed PRs, all together 3,720 2,715

On a Copilot PR, the requester’s own instructions to the agent count, and on a maintainer’s PR, the author’s own replies do, so both columns are every person who wrote on the PR. A merged Copilot PR took about twice as much of it. Some of that is steering (“@copilot now add similar optimization to the GetHashCode overrides”), which is work too: it’s just a different kind than review.

Agents people drive

The Copilot CLI and Claude Code run on the developer’s machine: they write the code with the person, and the person opens the PR (the table above). The Copilot CLI’s biggest gaps were in microsoft/aspire (23% against 11%) and PowerShell (28% against 8%, 72 PRs); in dotnet/runtime it was 16% against 14%. Claude Code’s 7 repos are mostly ones with a single main maintainer, such as SubtitleEdit, with 3,218 of its 4,133 PRs, so it says less about teams.

In these repos, handing a task to Copilot’s agent and coming back to a PR is where the PRs got thrown away, not AI as such.

What this can’t tell you

If you maintain one of these repos and something here looks wrong, tell us at [email protected] and we’ll correct it.

See it for yourself

The numbers for every repo are in repos.csv, one row each: PRs merged, closed and open, for Copilot’s agent, the Copilot CLI, Claude Code, maintainers and outside contributors. coderot runs as a CLI in your CI and your code never leaves it. Try the beta, free, to see your own repo.

Get the next report by email. We publish them as the data comes in, and now and then news about coderot.

Double opt-in. Unsubscribe anytime. How we handle your email.

See any public repo’s last 6 months, measured the same way: paste it here, no sign-in.