At last year's Build conference, GitHub announced a major update: Copilot can now work as an independent agent.
How to use it
It's simple — assign an issue to Copilot on GitHub, or give it instructions in VS Code. It will automatically start a secure development environment (based on GitHub Actions) and get to work:
- Read code and understand context
- Plan modification solutions
- Modify code across multiple files
- Run terminal commands
- Iterate automatically until completion
After finishing, it pushes the code to a draft PR. You can check the progress log at any time, and send it back for rework if you're not satisfied.
What it's good at
Currently, it performs well in these scenarios:
- Feature iteration (adding features on top of existing code)
- Bug fixes
- Test supplementation
- Code refactoring
- Documentation updates
In short, it works best on projects with low to medium complexity and good test coverage.
Safety mechanisms
A key design: PRs generated by the agent will not trigger any CI/CD pipeline until manually approved. Branch protection rules still apply. This means even if it writes problematic code, it won't automatically make it to production.
It became available to all paid Copilot users on September 25 last year. In March this year, MCP protocol support was added, officially upgrading it from a code completion tool to a full-featured agent development partner.
Each agent call consumes one premium request credit. For large projects, credit consumption needs to be monitored.
Source: CocoLoop, GitHub Blog official release