GitHub Launches New Coding Agent

Featured image: GitHub launches its coding agent

Introduction: The GitHub Copilot Coding Agent made its debut at the Microsoft Build conference held yesterday.

Earlier this year, GitHub previewed Project Padawan, an agent designed to assist software engineers and handle routine tasks such as code review, refactoring, and troubleshooting on their behalf.

Yesterday (May 19th), at the annual Microsoft Build conference, GitHub released the first version of this coding agent, named GitHub Copilot Coding Agent. This new Software Engineering (SWE) agent will be fully integrated into the GitHub experience and will be capable of receiving GitHub issues and handling them autonomously. It is launching today for Copilot Enterprise and Copilot Pro+ users.

Because the coding agent requires significant computing power, it will run in the cloud using GitHub Actions. Once a developer assigns an issue to the agent, it will spin up a customizable development environment in GitHub Actions to work on a pull request.

However, the agent will not be able to run any CI/CD workflows without human approval. It can only push to branches it creates itself and cannot touch default branches or other branches created by developers. Developers can limit the MCP servers and websites the agent can access. Additionally, to further streamline the review process, a developer who requests the agent to open a pull request cannot approve it themselves.

Image

“The trust factor of agents is not as high as the trust factor of human developers in today’s world,” GitHub CEO Thomas Dohmke told us. “It's almost like you wouldn't onboard someone to your team without doing an interview, without doing a background check, without doing all these things. So we think agents need to be in a more controlled environment than a human developer.”

Currently, it's powered by Anthropic’s Claude Sonnet 3.7 because, as Dohmke told me, the team thought the model currently had the “best combination of code quality and matching developer preference.”

Dohmke recently talked about Copilot becoming the developer’s peer. In many ways, that’s definitely happening.

“The workflow is: as a developer, you're still going to spend most of your time, hopefully on VS Code, building software that you love, doing what you love doing,” he said in an interview ahead of today’s announcement. “When you have a task, or somebody assigns a GitHub issue to you, or there's a bug report, you're going to throw it over the wall to the coding agent. And then you're going to have some agent sessions running in the cloud, and your local machine is available for you, being in flow. Hopefully doing magical things.”

Featured image: GitHub Copilot wants to be your peer programmer

Developers can assign an issue to the agent using GitHub.com, GitHub Mobile, or the GitHub CLI, just like they would to a colleague. The agent will then add an 👀 emoji and start working. Developers can monitor progress by viewing the agent’s reasoning steps and code validation efforts in the session logs.

GitHub stressed that all these features are best suited for well-defined requests that can typically be addressed with a single GitHub issue. “The agent excels at low to moderate complexity tasks in a well-tested code base from adding features, fixing bugs, to expanding tests, refactoring code and improving documentation,” Dohmke said.

From Code Completion to Coding Agents

Image

This new agent complements Copilot’s existing agent mode, which runs in the IDE. The agent mode can also write code from scratch, edit existing codebases, and use tools as needed. This agent mode, by the way, was previously only available in VS Code but is now also available for JetBrains, Eclipse, and Xcode.

Dohmke sees the current offerings as a continuous spectrum, from code completion and agent mode — which cover the inner loop of coding — to the coding agent, he stressed.

“The idea,” he explained, “is imagine the coding agent creates a pull request and has created five commits. It’s almost done, and now you need to make a decision: Is it going to be sustainable for you to keep prompting it until it gets exactly where you want it? Or is it maybe just faster for you to check out that pull request using the GitHub CLI? Open VS Code — with or without agent mode — make the change, and push it back? A pull request gives you an ideal point to say the agent got you this far, and now you do the last few commits back to the same pull request to get it ready for merge.”

Microsoft’s SRE Agent

In addition to the coding agent, Microsoft also launched a Visual Reliability Engineering (SRE) agent for Azure this week, which can actually use the new GitHub SWE agent to fix issues autonomously.

“The key thing here is that it can monitor your systems around the clock, and it can automatically troubleshoot issues that come up,” Amanda Silver, Microsoft’s Corporate Vice President and Head of Product for the Developer Division, told me. “And it can work together with your GitHub workflows. So once it figures out the root cause of an issue, it can actually automatically try to fix it, […], but it can also log these issues to GitHub. So we also have this really nice interaction where you can have the SRE agent monitoring your production systems. It finds an issue, it logs that issue as an issue to GitHub, and then the SWE agent, which is the GitHub Copilot coding agent that we're launching, can pick up that issue and go work on the fix.”

Silver noted that this SRE agent is based on an internal agent Microsoft developed and uses itself. Its data is based on the same guidance the company gives its engineers to help them troubleshoot Azure services.

Microsoft also partnered with New Relic to bring data from its application performance monitoring (APM) service into this workflow. The agent can also work with ServiceNow, PagerDuty, and other incident management systems.

Bringing the Fun Back to DevOps

All of this is about bringing the fun back to DevOps, Silver said.

“We think this whole category is actually ‘Agent DevOps,’ which is the next phase of DevOps as a whole,” Silver said. “This is the world where we see AI agents being embedded into every phase of development: from planning, to production, from coding, to deployment. This will really help to accelerate software delivery, increase quality, and bring back a lot of the fun.”

Bonus: VS Code Copilot Extension Is Now Open Source

The coding agent was undoubtedly the highlight of yesterday’s GitHub announcements. But it’s worth noting that GitHub is also open-sourcing the VS Code GitHub extension. VS Code is already open source, and so are many of the extensions developers rely on. The Copilot extension — including its system prompts — will now live in the same GitHub repository as VS Code.

“I think the key thing here is that this is really going to foster the ecosystem to co-evolve with us,” Silver told us. “VS Code has always been open source, and the extensions are where the magic of VS Code lies. I think, obviously, there's a ton of innovation happening in the space of AI-assisted coding, and so we want to make sure that VS Code continues to be at the center of all of that innovation.”

Notably for developers, the Copilot Extension for Apple’s Xcode IDE is already open source.

Editor: 场长

Related links:

Related reading:

Main Tag:AI Coding Agent

Sub Tags:GitHub CopilotMicrosoft BuildDevOpsSoftware Development


Previous:Google CEO Pichai Responds to "Google Is Dead"

Next:Nature Sub-journal: Humans Lost to AI Again, Especially When It Knows Who You Are

Share Short URL