GitHub Copilot Gets a Coding Agent: Autonomous Background Tasks and Pull Requests

At Microsoft Build 2025, GitHub introduced the Copilot coding agent: an asynchronous, autonomous developer agent. Hand it a task or issue and it runs in the background on GitHub Actions, submitting its work as a pull request (GitHub blog, 2025-05-19).

How it works#

As the official blog describes it: assign an issue to Copilot, or prompt it with `@github` in your editor, and the agent goes to work in the background — booting a virtual machine, cloning the repository, configuring the environment and analysing the codebase with retrieval-augmented generation (RAG) powered by GitHub code search. It pushes changes to a draft pull request as it goes, with reasoning and validation steps visible in the session logs. MCP servers can be configured in repository settings to give it outside data and capabilities, and with vision models it can also read images attached to an assigned issue.

The security defaults are in the blog too: the agent can only push to branches it created; whoever asked for the pull request cannot be the one to approve it; Actions workflows will not run without approval; and its internet access is limited by default to a customisable trusted list.

Generally available on September 25, 2025#

GitHub's official changelog states that the Copilot coding agent became generally available on September 25, 2025 for all paid Copilot subscribers; Copilot Business and Enterprise administrators have to enable it from the Policies page first (GitHub changelog, 2025-09-25). Besides assigning an issue, you can hand work over from the agents panel on GitHub pages, or via the Delegate to coding agent button in Visual Studio Code.

What it does, and where the line is#

The blog positions it for low-to-medium complexity tasks in well-tested codebases — adding features, fixing bugs, extending tests, refactoring, improving documentation — with pull requests reviewed and approved by a human. "Asynchronous delegation" therefore covers work that is already broken down, not the requirement a person still has to understand.