What Happened: GitHub’s Cloud-Based Agent Drafts PRs Autonomously GitHub announced a major new capability for Copilot: GitHub has unveiled a significant upgrade to Copilot during Microsoft Build 2025: a cloud-based coding agent capable of drafting and iterating on pull requests directly within GitHub. While previous Copilot experiences focused on in-editor assistance via VSCode, this new agent runs asynchronously in the cloud, leveraging an infrastructure similar to GitHub Actions. Microsoft Build 2025 Developers can now assign tasks directly via GitHub or VSCode: > @github Open a pull request to refactor this query generator into its own class > @github Open a pull request to refactor this query generator into its own class The Copilot agent handles tasks autonomously: creating branches, iterating on PRs based on code review comments, and updating commits until the work is accepted - all without touching protected branches. Crucially, existing CI pipelines, branch protections, and review workflows remain intact, ensuring “trust by design.” With the addition of MCP, developers can even grant the agent access to external tools and data by configuring servers directly in the repository settings. This abstracts us from the actual implementation and allows us to focus on describing the tasks. This aligns with Codex Agents and others providing compute for the agents to run and do the coding. It also resonated with the broader movement toward headless agents - those that run independently, complete tasks, and report back when work is ready for review. Community Reactions: Awe, Caution, Fears, and Questions. Personally, using GitHub Copilot Agent was an enjoyable experience when asking to perform low-level tasks. I had a ‘wow’ moment when I performed typical GitHub actions directly inside the PR (or via VSCode’s chat bar). It was impressively fast, and the developer experience felt spot on. When digging into the community’s reaction, we found Copilot’s coding agent showing a mix of excitement, curiosity, and healthy skepticism. “Time-saving” was the common praise. Users liked that the agent’s draft PR and log let them see exactly what was happening, and that they retained the choice of when/if to merge. Similar to my experience, users shared they found Copilot to perform well with low-level tasks. To which GitHub’s Product Lead for Copilot coding agent replied: A portion of the community also shared a fear that by offloading coding to AI, human work might devolve into janitorial oversight: filing detailed JIRA tickets all day and rubber-stamping AI-generated code. Also, users expressed concern about what this means for junior devs. The GitHub CEO’s take on this is: take In an age where some claim that more AI means fewer opportunities for entry level devs, I believe the opposite is true. There has never been a more exciting time to join our industry. In an age where some claim that more AI means fewer opportunities for entry level devs, I believe the opposite is true. There has never been a more exciting time to join our industry. In an age where some claim that more AI means fewer opportunities for entry level devs, I believe the opposite is true. There has never been a more exciting time to join our industry. Building on that concern, security and policy questions were also raised. For example, could the agent inadvertently expose sensitive information, such as including a secret key in a PR? Adding to the complexity, some developers pointed out that the benefits of using AI agents don’t come for free. You only truly reap the rewards if you invest effort upfront; e.g, writing a good issue description with clear acceptance criteria. Enjoy being informed about the AI Native Development space (AIND)? I’m helping build the AIND community, where we cover the most relevant dev-focused news. For the busy folks, we send out a crisp one-pager newsletter read by 7,000+ AI tinkerers every week - free and straight to your inbox! AIND community, a crisp one-pager newsletter by 7,000+ AI tinkerers every week The AIND Take: Abstracting Grunt Work and Developing New Habits GitHub is evolving into an AI-enhanced development platform. Much like the historic growth of Actions Marketplace, we can expect a spike in tools within Copilot Extensions, making Copilot more powerful and covering more aspects of the development workflow. Actions Marketplace Copilot Extensions Drawing an analogy from history, this feels akin to the DevOps revolution or the shift to cloud/serverless computing: mundane infrastructure work is abstracted away, enabling developers to focus on higher-level logic. Similarly, the Copilot agent can abstract away a chunk of grunt work. It’s also a new skill: prompting and supervising AI in coding. Just as CI/CD and cloud led to new roles, AI agents could lead to new roles like “AI orchestration engineer”. I suspect we’ll see AI-native startups experimenting more and more with having multiple agents collaborate. One use case we believe in is the use of task-dependent agents: one agent could generate code while another reviews it, or one agent specializes in front-end and another in back-end, coordinating through issues and PRs. GitHub’s infrastructure could support this kind of multi-agent workflow, especially as MCP allows chaining different AI services. From a developer’s perspective, to prepare for this AI-centric future, there are a few concrete steps we’d recommend when playing with Copilot’s Agent. First, writing clear, detailed task/issue/PR descriptions with acceptance criteria will become a valuable skill. It’s a good practice even for human collaborators, but for AI, it’s very relevant. This helps build the muscle of communicating with machines about code intent. very intent Second, invest in tests. The prospect of an AI agent that relies on tests to know it didn’t break things is a strong incentive. High test coverage can turn the agent into a reliable contributor, while low coverage turns it into a serious risk. Strengthening your automated tests and CI pipelines can help position your projects to benefit more fully from AI involvement. The end-game is to become comfortable letting the AI draft something, and you guiding/refining it. In this new era, the developers who thrive will be those who embrace specifications-centric development, guiding AI with precision, shaping code, and validating it with tests.
