GitHub Copilot Upgraded With Agent Mode and Gemini 2.0 Flash, Project Padawan Unveiled

GitHub Copilot, the artificial intelligence (AI) chatbot for the coding platform, is getting several new features. Announced on Thursday, the company is adding a new agent mode to the chatbot that can automatically iterate on code generated by it and analyse errors. Alongside, the Microsoft-owned platform also made Copilot Edits generally available, which was first announced at the GitHub Universe in October 2024. The company also took the curtains off Project Padawan, the platform's first software engineer (SWE) AI agent.

GitHub Is Introducing New Features to Copilot

Just a week after GitHub added the DeepSeek-R1 AI model to its platform, the company has made several improvements to the Copilot AI chatbot. As detailed in a blog post, the coding platform is now introducing a new agent mode to the chatbot.

The agent mode in GitHub Copilot will be more independent and proactive compared to the usual chatbot. It can iterate on its own code, recognise errors, and fix them without requiring intervention from a human user. It also comes with the ability to analyse run-time errors and it can fix them as well. Additionally, it will suggest terminal commands to users. Developers will be able to specify a task, and the agent mode Copilot will also complete all subtasks that might be necessary to complete the primary task.

Agent mode in Copilot can be activated by first downloading VS Code Insiders and then enabling the agent mode setting for GitHub Copilot Chat. After that, users will have to go to the Copilot Edits panel and switch from Edit to Agent via the model picker.

Copilot Edits is now generally available in VS Code. With the feature, developers can specify a set of files that require code rewrite or edits, and then using natural language, can describe what they want. Copilot Edits will contextually understand the command and then make inline changes to all the files. Users can review the suggested changes, accept them, and iterate on them with follow-up queries.

GitHub explained that the feature uses a dual-model architecture to improve editing efficiency and accuracy. A foundation language model, that a user can pick from GPT-4o, Claude 3.5 Sonnet, and Gemini 2.0 Flash, understands the full context of the Edits session and generates initial edit suggestions. Then, a speculating decoding endpoint, developed in-house, will propose those changes inline in the editor.

Finally, GitHub also shared the first look at Project Padawan, an autonomous software engineer (SWE) agent that will be shipped to Copilot later this year. The company said the AI agent can be assigned issues using any of the GitHub clients in Copilot, and it will produce fully tested pull requests. After finishing a task, the chatbot will also assign human reviewers to the pull request, and automatically resolve any feedback they leave.

As per the post, the AI agent will be provided with a secure cloud sandbox for every task it undertakes. Copilot oversees the entire process of adding the repository, setting up the environment, analysing the codebase, editing files and builds, and conducting tests.