GitHub is expanding Copilot's capabilities to work more independently alongside developers. The new Agent Mode transforms the AI assistant from a basic pair programmer into a more self-sufficient development partner.
Copilot's new Agent Mode can now spot and fix problems on its own, according to GitHub CEO Thomas Dohmke. The system suggests terminal commands and runs them after getting a thumbs-up from the developer. One of Agent Mode's key features is its ability to identify and handle additional tasks that weren't explicitly mentioned but are necessary to complete the main objective. The feature is currently available to VS Code Insiders users.
Speaking your code into existence
GitHub has also launched the full version of Copilot Edits, which lets developers modify multiple files at once using everyday language - including voice commands. The tool uses two models working together: one generates the suggested changes while another handles the editor integration.
Developers can choose their preferred AI model, whether it's GPT-4o, Claude 3.5 Sonnet, or the new Gemini 2.0 Flash. GitHub says it designed the tool to keep developers in control - they can review changes across files and pick what works. If something goes wrong, they can test the code and roll back to what worked before.
Training wheels come off
GitHub's Project Padawan offers an early look at what fully autonomous software development might look like. Set to launch later this year, this system will be able to handle issues and create tested pull requests on its own.
The tool will work directly within the GitHub interface, allowing developers to submit issues directly to Copilot. It's another attempt to make software development more autonomous, and potentially save companies a lot of time and money.
