TLDR: GitHub has introduced a new cloud-based coding agent for Copilot, enabling it to autonomously draft, iterate on, and manage pull requests directly within GitHub. This agent operates in a secure GitHub Actions environment, automating routine development tasks and freeing developers for more complex work. The feature was initially showcased at Microsoft Build 2025 and has seen continued enhancements and availability updates through October 2025.
GitHub has significantly advanced its AI-powered development tools with the introduction of a new cloud-based coding agent for Copilot. This innovative agent is designed to autonomously handle various aspects of the pull request (PR) workflow, marking a shift from in-editor assistance to a more independent, task-oriented AI teammate.
The Copilot coding agent operates asynchronously within a GitHub Actions-powered environment, allowing it to perform development tasks without direct, synchronous developer interaction. Its capabilities include creating new branches, writing detailed commit messages, pushing changes, opening draft pull requests, and even iterating on these PRs based on code review comments. It can also execute automated tests and linters to validate its work.
This new agent is distinct from the ‘agent mode’ available in Integrated Development Environments (IDEs). While IDE-based agent mode makes autonomous edits in a local development environment, the cloud coding agent works directly on GitHub, integrating seamlessly into the existing pull request workflow. This approach enhances transparency, as every step is recorded in commits and viewable in session logs, fostering better collaboration within development teams.
A key enabler for the agent’s expanded functionality is the Model Context Protocol (MCP). MCP allows the Copilot agent to access and leverage data and capabilities from outside of GitHub, with servers configurable in a repository’s settings. This means the agent can pull in relevant GitHub data and potentially interact with external services, expanding its problem-solving scope.
The primary benefit for developers is the offloading of time-consuming, low-to-medium complexity tasks such as refactoring, bug fixes, extending tests, and improving documentation. By automating these routine chores, the agent allows human developers to concentrate on more intricate, creative, or high-value work.
GitHub has implemented several safeguards to mitigate the inherent risks of an autonomous agent having access to code. These include operating with limited network access by default, requiring explicit human approval for actions and workflows, and ensuring that the individual who assigns a task to Copilot cannot also approve the resulting pull request. Existing security policies, such as branch protections, remain fully applicable.
Developers can delegate tasks to the Copilot coding agent by assigning GitHub issues to it, prompting it through GitHub Copilot Chat, or using the GitHub CLI. The agent then boots a virtual machine, clones the repository, configures the environment, and analyzes the codebase using advanced retrieval augmented generation (RAG) powered by GitHub code search. It can even process images included in GitHub issues, thanks to vision models, allowing for visual bug reports or mockups.
As of June 4, 2025, using the Copilot coding agent incurs one premium request per model request made by the agent. This pricing model applies across various Copilot features, reinforcing GitHub’s mission to keep developers in a ‘flow state’ by automating undesirable tasks.
Also Read:
- Enhancing Trust and Security for AI Agents in Cloud-Native Environments with Docker’s MCP and cagent
- Opsera Launches Next-Generation AI-Powered DevOps Platform with Hummingbird AI for Enhanced Code Quality Analysis and GitHub Integration
Further updates in October 2025 indicate ongoing development and integration, including the ability to kick off and track sessions from the GitHub CLI, context retention within the same pull request, and public preview for assigning Azure Boards work items to the agent.
