GitHub Copilot launches agent mode with self-healing code

The most annoying part of coding with AI isn’t generating the code—it’s acting as an unpaid human copy-paster. The AI gives you a script, you paste it into the terminal, it throws a wall of red text, and you paste that error right back into the chat to ask what happened.

GitHub is finally cutting out the middleman. The company just rolled out a major update headlined by an experimental “agent mode” for GitHub Copilot in VS Code.

Alongside the agent upgrade, GitHub also made its multi-file editing tool, Copilot Edits, generally available, added Google’s Gemini 2.0 Flash to the model picker, and teased an autonomous SWE agent codenamed Project Padawan.

Self-healing code inside your editor

Agent mode pushes Copilot from an autocomplete assistant toward an active collaborator.

Available now in VS Code Insiders, agent mode doesn't just stop at generating a code snippet. It inspects the actual output of its actions and iterates until the entire chain of subtasks is finished. If you don't explicitly spell out every requirement, Copilot can infer necessary steps, suggest terminal commands, and ask for permission to run them.

Crucially, it handles errors on its own. When runtime issues pop up in the terminal, Copilot analyzes the errors and applies fixes without forcing you to copy and paste back and forth. You trigger it in the Copilot Edits panel by switching the drop-down from “Edit” to “Agent.” GitHub says it plans to expand the capability across all supported IDEs.

Fast multi-file edits and autonomous PRs

Meanwhile, Copilot Edits is now generally available in VS Code and in preview for Visual Studio 2022.

The feature allows developers to point Copilot at a specific set of files and make conversational, inline changes across multiple files at once. Under the hood, GitHub uses a dual-model approach: a foundation model (now supporting OpenAI’s GPT-4o, o1, o3-mini, Anthropic’s Claude 3.5 Sonnet, and Gemini 2.0 Flash) sets the intent, while a speculative decoding endpoint rapidly applies the diffs directly into the editor.

Then there’s Project Padawan, GitHub's upcoming autonomous SWE agent expected later this year. Instead of living solely in your editor, Padawan will integrate directly into repository workflows. You will be able to assign an Issue directly to Copilot from any GitHub client.

Once assigned, Copilot spins up an isolated cloud sandbox, clones the repo, builds the environment, writes the code, runs tests, and submits a PR. From there, it assigns human reviewers and will iterate on whatever feedback they leave on the pull request.

Why it matters

The AI coding battle has officially moved past inline autocomplete into autonomous loops.

For individual developers, agent mode tackles the most tedious cycle in modern software engineering: running into dependency snags, reading stack traces, and fixing minor typos. Letting the model catch its own terminal errors saves real cognitive load, making the Insiders build worth checking out immediately.

Strategically, we think GitHub is playing its home-field advantage brilliantly. Third-party AI code editors like Cursor have gained massive momentum by offering slick multi-file edits and terminal awareness. GitHub isn't just catching up inside VS Code with speculative decoding and multi-model flexibility; it's leveraging the core GitHub platform.

Project Padawan represents the real end-game. By moving the agent into secure cloud sandboxes tied directly to Issues and PRs, GitHub can turn AI into an asynchronous background teammate. While other tools fight for real estate on your desktop, GitHub is positioning Copilot as a permanent contributor to your entire codebase.