GitHub Copilot coding agent enters public preview

Background

AI coding assistants were first used in a fairly narrow way: a developer typed in an editor, and the tool offered completions or snippets that the developer accepted or rejected. Deciding what to build, breaking it into steps and checking the result stayed with the person. The Copilot coding agent moved that boundary. The unit of work being handed over was no longer a suggestion inside a file but a task described in an issue.

An issue as the interface

On 19 May 2025 GitHub put the Copilot coding agent into public preview ^[1]. Starting work is deliberately ordinary: an issue is assigned to Copilot exactly as it would be assigned to another developer, from github.com, GitHub Mobile or the GitHub CLI ^[1]. There is no separate prompt tool to learn. The issue tracker, which teams already use to describe work, becomes the place where the request is made.

How a task is carried out

Once assigned, Copilot works in the background in its own secure cloud-based development environment powered by GitHub Actions ^[1]. Inside that environment it explores the repository, makes changes, and validates its work with the project's tests and linter before pushing ^[1]. Only then does it hand back: when finished, it tags the user for review, and further changes can be requested by commenting on the pull request ^[1]. Several issues can be assigned to Copilot at the same time, so more than one task can be under way at once ^[1].

What it is aimed at

GitHub positions the agent for low-to-medium complexity tasks in well-tested codebases: features, bug fixes, tests, refactoring and documentation ^[1]. Two conditions in that description carry weight. The tasks are small enough to be described in an issue, and the codebase has tests and a linter, which is what lets the agent check its own work before a human looks at it.

Who could use it, and what it costs

At launch it was available to Copilot Pro+ and Copilot Enterprise subscribers, and administrators on Enterprise had to enable a Copilot coding agent policy first ^[1]. Usage consumes GitHub Actions minutes and Copilot premium requests ^[1]. From 4 June 2025 the agent used one premium request per model request it made ^[1]. Availability also rolled out to GitHub Mobile on iOS and Android and to the GitHub CLI during that week ^[1].

Why it matters

The change is in where the person sits. Instead of watching code appear, a developer writes an issue, waits, and then reviews a pull request, asking for revisions by commenting on it ^[1]. That makes review the main point of contact between human and agent, and it puts weight on the project's own tests, since they are the check the agent runs before anyone reads its work ^[1]. It also makes cost visible in two currencies at once: Actions minutes for the machine work and premium requests for the model calls ^[1]. Because the agent is intended for codebases that are already well tested, the tasks it suits best are the ones a team can describe precisely and verify automatically.