GitHub has announced new features for GitHub Copilot to streamline coding tasks. Copilot will now predict what a developer might be going to type next based on what they have already typed. The tool also gets the ability for its agents to implement changes across multiple files.

Copilot is GitHub's AI-based code helper system. It was originally based on OpenAI's Codex, but recently added multi-model support meaning developers can select from multiple models. In addition to Anthropic Claude Sonnet, Google Gemini Pro and OpenAI GPT, this latest update adds support for Gemini 2.0 Flash.

The updates that have been announced aim at streamlining coding tasks based on an organization’s specific ways of working. These improvements start with Copilot Edits, which let a developer specify a set of files to be edited and then use natural language to ask Copilot what changes should be made. Copilot Edits makes inline changes directly in the user's workspace, across multiple files. If that sounds like a disaster waiting to happen, GitHub CEO Thomas Dohmke says:

"You stay in the flow of your code while reviewing the suggested changes, accepting what works, and iterating with follow-up asks."

Behind the scenes, Copilot Edits uses a dual-model architecture. A foundation language model (chosen by the developer) considers a full context of the Edits session to generate initial edit suggestions. These suggestions are then passed to a "speculative decoding endpoint, optimized for fast application of changes in files" that has been developed by GitHub. This is then used to propose changes inline in the editor. Dohmke said in a blog pos that Copilot Edits works because it puts you in control, from setting the right context to accepting changes. When the model gets it wrong, you can review changes across multiple files, accept good ones and iterate until, together with Copilot, you arrive at the right solution.

After accepting changes, developers can run the code to verify the changes and, when needed, undo in Copilot Edits to get back to a previous working state. The new agent mode will also recognize errors and fix them automatically.

In agent mode, Copilot will iterate on not just its own output, but the result of that output. And it will iterate until it has completed all the subtasks required to complete your prompt, inferring any additional tasks that were not specified, but are also necessary for the primary request to work.

GitHub has also provided a first look at a new autonomous SWE agent. SWE agents are a type of automated system that act on behalf of developers, performing tasks such as generating and reviewing code, refactoring or optimizing the codebase, automating workflows like tests or pipelines, and providing guidance. Dohmke said that when the product GitHub is building under the codename Project Padawan ships later this year,

"it will allow you to directly assign issues to GitHub Copilot, using any of the GitHub clients, and have it produce fully tested pull requests. Once a task is finished, Copilot will assign human reviewers to the PR, and work to resolve feedback they add. In a sense, it will be like onboarding Copilot as a contributor to every repository on GitHub."

The final element announced by GitHub is Vision for Copilot, which lets users bring a mock up to life by feeding Copilot a snip, screenshot, or image. From there, Copilot generates the UI, alt text, and code.

GitHub Copilot updates are available now.

More Information

Agent mode for GitHub Copilot in VS Code

Related Articles

The Github Copilot Mega Thread

GitHub Announces Free Copilot

Copilot Improves Code Quality

VS Code Extensions Can Now Use Copilot APIs

GitHub Universe AI Announcements - Copilot And Spark

Copilot Chat Improves Confidence and Enjoyment

GitHub Copilot Provides Productivity Boost

GitHub Enterprise Server 3.10 Improves Security

To be informed about new articles on I Programmer, sign up for our weekly newsletter, subscribe to the RSS feed and follow us on Twitter, Facebook or Linkedin.

Comments

Make a Comment or View Existing Comments Using Disqus

or email your comment to: comments@i-programmer.info