GitHub Copilot's new coding agent turns issues into pull requests

The most exhausting part of software engineering usually isn't building novel architecture. It's the mountain of chores that comes with it: updating tests, patching low-priority edge cases, and cleaning up documentation. GitHub is now trying to take that grunt work off your plate entirely.

GitHub Copilot just rolled out a new autonomous coding agent designed to live directly inside your existing repo workflow. Instead of pairing with you line by line inside an editor, this agent acts like an asynchronous junior engineer: you assign it an issue, and it comes back with a pull request ready for review.

From issue assignment to a draft pull request

The trigger is intentionally mundane. You can assign a GitHub issue to Copilot on the web, inside GitHub Mobile, or via the GitHub CLI, exactly the way you would assign work to a teammate. You can also trigger it straight from VS Code or Copilot Chat by prompting it to open a PR for a specific task.

Once assigned, the agent drops an eye emoji on the issue and gets to work in the background. It spins up an isolated virtual machine powered by GitHub Actions, clones the repo, configures dependencies, and indexes the project using retrieval-augmented generation tied to GitHub code search.

As it writes code, the agent pushes incremental git commits to a draft pull request and updates the PR description. Every reasoning step and validation check is recorded in transparent session logs, allowing engineers to follow along with its thought process.

It doesn't just read plain text, either. Thanks to vision models, you can paste UI mockups or bug screenshots directly into the issue. Furthermore, with Model Context Protocol (MCP) support, teams can plug in external context from outside GitHub alongside the official GitHub MCP server.

When it finishes, Copilot pings you for a review. Leave a comment pointing out an error, and it picks up the feedback to push new commits automatically.

Guardrails and the Actions compute engine

Handing code generation keys to an autonomous bot naturally raises security alarms. GitHub built in several hard guardrails by default.

The agent can only push to its own newly created branches. The developer who requested the PR is barred from approving it, ensuring that existing branch protection rules and peer reviews remain mandatory. Crucially, CI/CD pipelines will not execute on the agent's pull requests without explicit human sign-off, stopping bad code from quietly entering the build pipeline. Its outbound internet access is also locked down to customizable trusted destinations.

Under the hood, GitHub built this on GitHub Actions, which already executes over 40 million jobs every weekday. In terms of pricing, GitHub noted that beginning June 4, 2025, each model call made by the coding agent will consume one premium request.

Why it matters

For developers, this shifts AI from an autocomplete gimmick to genuine asynchronous delegation. Enterprises like EY and Carvana are already using it to offload repetitive tasks so senior engineers can stay focused on core architecture. If your codebase has strong test coverage and clear documentation, this is worth testing immediately on low-to-medium complexity tickets.

Strategically, this is GitHub flexing its moat against standalone AI coding startups. While third-party agents struggle to build reliable remote sandboxes, GitHub already runs the largest CI/CD compute footprint on earth with GitHub Actions. Integrating agentic execution straight into issues and PRs creates zero adoption friction for existing enterprise teams.

The catch lies in complexity and cost. Autonomous agents excel when tasks are well-defined and tests are comprehensive, but they quickly degrade in messy, legacy repositories. With requests burning premium quota starting in June, teams will have to keep a close eye on whether the agent actually solves problems on the first try or burns compute spinning its wheels.