GitHub Copilot coding agent

GitHub Copilot's coding agent, which GitHub's documentation now calls the Copilot cloud agent, takes a task, works in the background and opens a pull request for a person to review. GitHub announced it in public preview on 19 May 2025 for Copilot Pro+ and Copilot Enterprise. It became generally available to all paid Copilot plans on 25 September 2025, with an administrator switching it on for Business and Enterprise accounts.

We read 12 sources: seven official pages from GitHub and Microsoft, three news pieces, one independent write-up and one Hacker News thread. They were published from 19 May 2025 to 26 May 2026, and we read two pages and ran one public search on 10 October 2026. The independent hands-on tier is empty: two trade-press hands-on articles returned HTTP 403 and could not be read. The announcement promised a pull request. This review holds that promise against what developers report about merging them.

Consensus

Across the sources we read, the agent does well on narrow, well-specified work, and its pull requests are drafts for a person to approve rather than finished changes. Five of the 12 sources describe a person reviewing before anything merges as part of the design: three official pages (GitHub's preview and general availability announcements, and the .NET team's May 2026 advice) and two news pieces (InfoWorld and Hackernoon).

The announcement said the agent suits low-to-medium complexity tasks in well-tested code. The later reports agree with that scope. Where they pull away from the announcement is review: the launch text says the agent tags you for review, and none of the sources we read sizes how much review that creates. We read this as an announcement that was accurate about what the agent attempts and quiet about the cost on the reviewer's side. That is our interpretation, not a finding any source states.

Confidence is mixed to thin. The official evidence is detailed but comes from the makers and their parent company. The independent evidence is three news pieces, one write-up and one forum thread, and nothing from owners after months of use that we could read.

Recurring strengths

Six of the 12 sources say the agent is best on bounded, clearly described tasks: three official (GitHub's preview announcement and the .NET team's two posts), two news (InfoWorld and Hackernoon) and one independent write-up (Start Debugging). The .NET team's ten-month report, covering 19 May 2025 to 22 March 2026 in the dotnet/runtime repository, gives its success rate by task: 84.7% for removal and cleanup, 75.6% for tests, and 69.4% for bug fixes.

Two news sources, InfoWorld and Hackernoon, say the draft pull request and session logs make the agent's work easy to follow. One official source, the .NET report, says setup mattered most: its success rate rose from 38.1% to 69% once the repository held a written instructions file with build commands and test patterns. That is one team's account of one repository.

Recurring complaints

Four of the 12 sources raise review capacity: one independent write-up (Start Debugging), one forum thread (Hacker News) and two news pieces (Hackernoon and Let's Data Science). Start Debugging reads the .NET report as saying one developer can create five to nine hours of review work remotely. Hackernoon reports unease that human work becomes approving machine output. The Hacker News thread, from launch week, doubts that busy teams will scrutinise every change. Let's Data Science writes about agent-written pull requests in general, not this agent, and warns that tidy-looking changes can hide duplicated logic. No source counts how many teams are affected.

Three sources, two official and one independent, show that not every pull request lands. In the .NET report, 535 of 878 pull requests were merged. On 10 October 2026, GitHub's public search reported 1,528,432 merged out of 2,118,197 pull requests opened by the agent's bot account, a count that includes pull requests still open.

Three sources, one official, one independent and the Hacker News thread, describe limits on harder work. The .NET report gives 54.5% for performance optimisation, says the agent runs on Linux only so it cannot check Windows or macOS code, and says pull requests with a person's own commits on top reached 86% against 55% for fully autonomous ones. Developers in the launch-week thread said the code works but is structured poorly. We did not count how many.

Where reviewers split

How much gets merged. The .NET report counts one repository. GitHub's search counts every public pull request from the bot account. Neither says how carefully the merged ones were read, and a merge count is not a quality measure.

What setup is worth. The .NET post puts the success rate at 38.1% before an instructions file and 69% after. Start Debugging, writing about the same report, gives 41.7% and 72%. We use the report's own figures, and treat the difference as a reason to read the original.

Announcement against the thread. InfoWorld reports that every pull request needs human approval before CI workflows run. The launch-week Hacker News thread asks whether people will do more than approve. Both can be true, since the first describes a gate and the second describes habit.

Then and now. The preview announcement named two plans, and the general availability announcement names every paid plan. GitHub's changelog of 5 August 2025 says the agent now answers only to people with write access who mention it, and the documentation read on 10 October 2026 limits a session to 59 minutes and one pull request per task.

Who it suits

You maintain a GitHub-hosted repository with working tests and can write down build and test commands for the agent. You have a backlog of cleanup, test additions and well-described bug fixes, and someone with time to read each pull request.

Who should pass

You cannot spare reviewer time, or you need Windows or macOS-specific changes checked before merge. Performance tuning and design decisions are also where the .NET report records its weakest results. If your code lives outside GitHub, the documentation says the agent does not work there.

One verdict a week: the most useful Master Review we finished, the complaint that kept appearing, and who should skip it.

Sources

• GitHub changelog, Copilot coding agent in public preview, official, 19 May 2025
• InfoWorld, GitHub unveils coding agent for GitHub Copilot, news, 20 May 2025
• Hacker News, launch-week thread on the coding agent, forum, 19 May 2025
• GitHub changelog, improved pull request review experience, official, 5 August 2025
• GitHub changelog, coding agent generally available, official, 25 September 2025
• Hackernoon, GitHub's Copilot adds cloud agent to draft pull requests, news, 26 October 2025
• .NET blog, Ten months with Copilot coding agent in dotnet/runtime, official, 23 March 2026
• Start Debugging, the dotnet/runtime ten-month data, independent write-up, 29 March 2026
• Let's Data Science, engineers review agent-generated pull requests, news, 7 May 2026
• .NET blog, doing more with GitHub Copilot, official, 26 May 2026
• GitHub Docs, about the Copilot coding agent, official, read 10 October 2026
• GitHub search API, pull requests by the agent's bot account, official, searched 10 October 2026

On 10 October 2026 we read 12 sources from 19 May 2025 to 26 May 2026 by web search and plain page fetch, and synthesised them with AI; we collected no developer reviews, ran nothing and tested nothing.