GitHub Copilot Upgraded with Agent Mode, Gemini 2.0, and Project Padawan

Consultants Review Team

GitHub Copilot, the AI chatbot for the coding platform, is receiving a range of new features. Announced on Thursday, the company is introducing a new agent mode that automatically iterates on the code it generates and analyzes errors. Additionally, GitHub has made Copilot Edits widely available, a feature first revealed at GitHub Universe in October 2024. The company also unveiled Project Padawan, its first AI agent designed for software engineering (SWE).

GitHub Unveils New Features for Copilot

Just a week after adding the DeepSeek-R1 AI model to its platform, GitHub has introduced several enhancements to the Copilot AI chatbot. As outlined in a blog post, the coding platform is launching a new agent mode for the chatbot.

The agent mode in GitHub Copilot is designed to be more autonomous and proactive than the standard chatbot. It can independently iterate on its own code, detect errors, and fix them without needing human intervention. Additionally, it can analyze run-time errors and resolve them. The feature also includes the ability to suggest terminal commands to users. Developers can specify a task, and the agent mode Copilot will handle all related subtasks required to complete the main task.

To activate Agent mode in Copilot, users must first download VS Code Insiders and enable the agent mode setting for GitHub Copilot Chat. After that, they can go to the Copilot Edits panel and switch from Edit to Agent using the model picker.

Copilot Edits is now generally available in VS Code. This feature allows developers to specify a set of files that need code rewrites or edits, then use natural language to describe the desired changes. Copilot Edits will understand the command context and make inline updates to all the files. Users can review the suggested changes, accept them, and refine them through follow-up queries.

GitHub explained that the feature utilizes a dual-model architecture to enhance editing efficiency and accuracy. A foundational language model, which users can choose from GPT-4, Claude 3.5 Sonnet, and Gemini 2.0 Flash, comprehends the full context of the Edits session and generates the initial edit suggestions. Then, an in-house developed speculating decoding endpoint proposes those changes inline within the editor.

Lastly, GitHub revealed a preview of Project Padawan, an autonomous software engineer (SWE) agent set to be integrated into Copilot later this year. The company stated that the AI agent can be assigned issues through any of the GitHub clients in Copilot and will generate fully tested pull requests. Once a task is completed, the chatbot will assign human reviewers to the pull request and automatically address any feedback they provide.