GitHub Bets AI Can Bring More People Into Coding While Pushing Pros Further
GitHub logo on display at Microsoft Build 2025 in Seattle, Wash. Credit: Ken Yeung

Somewhere in prehistory, someone figured out how to make fire, showed others, and watched the knowledge spread until anyone could summon a flame. Mario Rodriguez, GitHub’s chief product officer, considers it probably the first open source project on Earth, and he sees that same instinct to share and remix at the heart of everything his company has built.

That instinct showed on GitHub’s homepage in 2008, which listed individual developers alongside their public repositories, all beneath a “social coding” tagline. Over time, the site became the place where developers and eventually organizations published their work so others could pick it up (or “fork”) and improve on it. Rodriguez insists that mission still holds, and he believes the community now includes not only human developers but also AI agents.

This expanded community frames the announcements GitHub is making at its Universe conference in San Francisco this week. The new features, from the launch of Copilot Fusion to enhancements to the Copilot app, are designed to help developers work alongside AI, directly addressing fears that the technology will replace them. Rodriguez described the company’s goal as lowering the floor so far more people can create software, while simultaneously raising the ceiling for professional developers who want to push their craft further.

‘Lower the Floor and Raise the Ceiling’

Rodriguez sees AI as the first technology capable of opening software creation to far more people while expanding what experts can accomplish. By lowering the floor, he expects the number of people building software to grow dramatically, from roughly 200 million to a billion. At the same time, he argues that AI is handing professional developers new tools that could speed up breakthroughs in fields like cancer research. And no matter who the builder is, Rodriguez wants GitHub to be the place where it all happens.

“We’re going to be obsessed about completed software work, and allowing the developer to delegate that, and then creating new tools that then take craft to a new level,” he explained. “This is why I’m saying raising the ceiling.” He believes the right tools will let the industry’s top developers “leverage 10x, 100x, 1,000x” their abilities. For him, it’s a special time because “we never had this where you actually can take a technology, lower the floor and raise the ceiling at an accelerated rate, and with the availability and accessibility that we have today.”

The wave of new builders Rodriguez envisions is quietly redefining what it means to be a GitHub user. The platform once served mostly trained developers, but Rodriguez now pictures it as a home for anyone who builds, whether they’re vibe coding a personal website or engineering software at the frontier of their field. The nature of collaboration is changing too. GitHub was built to connect developers sharing code with one another, and its users now also work alongside AI agents, directing them to plan, write, and fix code on their behalf.

Intelligence, Velocity, and Quality

Rodriguez grouped the GitHub Universe announcements into three themes. Intelligence covers how Copilot thinks, led by the launch of Copilot Fusion and new capabilities that let Copilot learn a developer’s preferences and codebase over time. Velocity focuses on helping developers get more done, with Copilot app updates that let them hand off routine work to AI agents and track it all from one place. Quality addresses the risks that come with that speed through AI-powered security scanning and upgraded code review designed to catch serious bugs before they reach customers.

“The [Software Development Lifecycle] is no longer linear,” writes Jay Parikh, Microsoft’s executive vice president for CoreAI, which oversees GitHub, in a blog post. “Often, a signal or an event like a code change, request from a colleague, or a spike in a production metric triggers actions in GitHub core, which then produce outcomes in the form of completed work. Most of us jump between stages, depending on the task at hand. Agents handle the bulk of the implementation and verification work, and everything feeds into and out of a continuous learning machine in the center.”

“We can now build a truly agentic software delivery system that solves for velocity, intelligence, and quality,” he adds.

Copilot Fusion

This is the centerpiece of all GitHub Universe announcements. It’s a system inside GitHub Copilot that manages which AI models handle a developer’s work as it happens. Copilot Fusion combines multi-model orchestration from Project HydraFusion with fine-tuned model selection in Copilot’s auto model selection feature. It evaluates what a developer is asking for and decides which AI model best handles each task, choosing from one of three initial algorithms. With “Single,” Copilot Fusion picks one model as the right fit for the task and uses it from start to finish. Next is “Cascade,” where one model plans and another executes. Lastly, “Critique” has one model propose a plan, a second review and challenge it, and the two debate before completing the work. Rodriguez said more algorithms are planned for the future.

Credit: GitHub
Credit: GitHub

Copilot Fusion is now generally available and comes in three levels depending on how much effort a task requires. Fusion Standard prioritizes versatile models for everyday coding, Fusion High is for challenging tasks and deeper reasoning, and Fusion Max is for the most demanding tasks and complex reasoning.

When tested using the DeepSWE benchmark with the GitHub Copilot harness, Fusion High and Max generated task-completion scores up to 8.7 points higher than similar efforts from leading frontier models. They also reduced estimated costs by up to 77 percent. GitHub also evaluated Fusion Max against the Artificial Analysis Coding Agent Index, where it scored 60, two points below Anthropic’s Claude Code with Opus 5.5 Max. GitHub says Fusion Max delivered that effort with half the average API cost per task, with half as many tokens and 40 percent less time per task. Rodriguez frames the payoff simply: helping developers do more by making their tokens go further.

He said that Copilot Fusion can continue working when a developer goes offline. Rodriguez explained that it can hand tasks to an AI model running on the developer’s own device, so Copilot remains useful offline, for example on a plane without Wi-Fi. In that situation, Copilot Fusion relies on the Single algorithm because it can’t reach models in the cloud. When developers are connected, they get what Rodriguez calls “the best of both worlds”: a powerful cloud model plans and coordinates the work while their device handles some tasks. He argued that this combination delivers the real payoff because it lets developers delegate work without “spending all of these token costs,” the usage fees that add up every time AI does work on their behalf.

YouTube player

Copilot Fusion will only use models that are enabled in the enterprise and organization model policies. Administrators can control which models can be chosen, set each model’s effective policy state independently, and even disable models for Copilot Fusion without disabling them for everyone. GitHub said customers will be charged for each model Copilot Fusion uses at each model’s standard rate. However, for a limited time, paid Copilot subscribers will receive a 30 percent discount when using Copilot Fusion.

All Copilot plans have access to Copilot Fusion. It can be used in Visual Studio Code, the GitHub Copilot app, GitHub Copilot command line interface, and JetBrains’ integrated development environment (IDE). Support for more clients is expected in the future.

Yet GitHub isn’t the only platform to adopt model routing. Once companies grew tired of tokenmaxxing, they sought out solutions to make their AI usage more efficient. And it’s a crowded marketplace—along with Copilot Fusion, there are tools from Snowflake, Amazon Web Services, Databricks, Stripe’s OpenRouter, Palo Alto Networks via Portkey, TrustedRouter, Salesforce, and also GitHub’s parent company, Microsoft.

Copilot App Updates

Credit: GitHub
Credit: GitHub

GitHub is also adding new features to its Copilot app that make it easier to coordinate projects, agents, and delegated work across tools and devices. It’s the first major update since the company launched it in May 2026. Rodriguez described the changes as a way to reduce a developer’s cognitive load and keep them focused on building. “What we want to do is keep you in the flow as much as possible,” he said. “No developer wakes up and is like, ‘Let me do ten dependency updates today.’ No developer wakes up and is like, ‘Oh my God! There’s a customer with this simple bug. Let me go and spend time on that at the moment.’ They just want to delegate a lot of these things.”

Copilot Projects:

YouTube player

This feature centralizes everything related to a large piece of work, from the set goals to the progress AI agents have made. It includes an orchestrator that lets developers manage multiple sessions at once and monitor status from a single screen. 

“As you create more and more and more, you end up in a place where you have to have all the cognitive load of what your agents are doing. And guess what? An orchestrator agent can actually do a really good job at that, and only bring you in the places where you need to actually make a set of decisions that keeps you in the flow overall,” Rodriguez said.

Copilot Connectors:

GitHub described this tool as helping simplify how Copilot connects with MCP servers like those from Vercel, Cloudflare, and Sentry. Copilot Connectors work wherever developers use Copilot, whether through the app, the command line interface, VS Code, or Copilot code review.

Copilot Automation:

YouTube player

This is where developers can find reusable agent tasks, specify where they run, and monitor costs. Rodriguez sees the feature as a way to protect the creative side of software development, likening it to music. “If you think of a jazz player playing and being in the flow…you’re never going to automate that in your life,” he explained. Automations instead take on the chores that interrupt that flow. “What we really care about, though, is…delegation of the things that are actually inhibiting the flow of these developers,” Rodriguez argued. 

He added that developers can set up automation “squads” for specific kinds of work, each bringing relevant expertise. This frees them up to focus more on creativity and “actually making the idea in your head possible to the world.”

Cloud Sandboxes:

Software development can tax a computer, especially when AI agents are running in the background. Rodriguez knows the feeling, joking that whenever he compiles Rust code, his computer sounds like a helicopter. GitHub’s answer is cloud sandboxes, which move the agents’ work to the cloud so it doesn’t consume all the processing power and memory on the developer’s local machine.

Continuity:

GitHub wants developers to be able to work on projects wherever and on whatever machine or device they want. It’s why the company introduced a continuity feature. As Rodriguez explained, the feature lets developers start a project on their laptop and then switch to their mobile device to keep working. They could then carry on in VS Code or even Slack. Continuity makes it easier for developers to keep building.

Better Code Quality and Security

The final set of announcements focuses on improving developers’ code, helping them identify problems faster, mitigate risks, and patch vulnerabilities. Rodriguez ties the effort to the global economy, arguing that if AI-assisted software is going to boost gross domestic product, that growth will come through businesses, and those businesses need to ship secure, working code. “Upholding the engineering bar is critical, and even more critical today when we’re lowering the floor of [actual] creation as well,” he stressed. “We need the whole industry to understand that upholding the engineering bar is critical for us to be successful.”

Copilot Code Review:

GitHub made Copilot code review generally available in April 2025. Developers can use the tool to offload basic reviews to a Copilot agent, tasking it with spotting bugs and potential performance problems and recommending fixes. The company is rolling out updates for it at Universe.

The first is a new max tier, separate from Fusion Max. GitHub said this premium option will catch more complex bugs before they reach production, though it comes with a higher price tag. 

The second update gives code review a memory. Rodriguez explained that Copilot as a whole can now “dream” during sessions, updating its instructions, skills, and memory to better match the developer and their codebase. Because code review draws on that same memory, Rodriguez expects it to get “better and better” over time.

As developers spend more time reviewing code, GitHub reasons that some changes don’t need a human to look at them. That’s why the company is introducing an auto-approval feature and says it’s developing algorithms that not only identify low-risk changes, but also tell developers how risky each one is to the application they’re building, letting them approve it automatically. Developers can review the change later, after it passes automated tests and reaches production.

AI Scans:

GitHub says this feature analyzes every pull request to detect vulnerabilities in languages that rule-based scanning typically misses. It also scans repositories to check vulnerabilities in already merged code.

Reachability Analysis:

Security alerts often flag a flaw in outside code a project relies on, even when the project never uses the flawed part. Reachability analysis checks whether a developer’s code actually reaches the vulnerable piece and shows the path it takes, helping teams focus on the risks that matter. The capability is also coming in preview to Dependabot, GitHub’s tool for flagging outdated or vulnerable code libraries.

Package Firewall:

GitHub can now check packages from npm, a popular JavaScript package registry, against an organization’s enterprise policy before they’re installed, blocking known malware and critical vulnerabilities before they enter a build.

GitHub Enterprise Update:

And there’s one more thing from GitHub. Enterprise customers can now use all GitHub Copilot features without needing a separate subscription. Under a single Copilot license, customers can use GitHub Copilot app, Code Review, Automations, and more—they only need to pay for AI credits. This change is available today through GitHub Sales, but a self-serve option is coming soon. Current Copilot Business and Enterprise customers are grandfathered in so they won’t need to update their plan, at least for now.

Agent-Proofing GitHub

Despite these updates and the message Rodriguez wants to send, developers may remain skeptical about GitHub’s future. The community is still wary of the platform’s performance after multiple outages earlier this year. The most severe one, in August, lasted nearly eight hours and took GitHub Copilot down with it. While GitHub sees AI agents benefiting developers, those same agents are straining its infrastructure. By the company’s own count, monthly commits more than doubled between April and August, to 2.9 billion.

“We are the home of developers, which means…we have a responsibility to be able to take that load,” Rodriguez acknowledged. To carry it, GitHub added four million CPU cores, 120 petabytes of high-speed storage, and 50 terabytes per second of network capacity this year to shore up its systems. 

Another effort to stabilize the platform is the move to Azure, which GitHub committed to fully in 2025. As of August, Microsoft’s cloud service handled roughly 58 percent of GitHub’s platform load, up from 12 percent in May. “We ran out of power in the data centers that GitHub was running in, so we’re doing this migration to Azure to really scale to that 100x that the world needs us in,” Rodriguez shared. “That’s the journey that we’re in right now. It’s like, the agents and all these developers can continue to create, and we’re going to go and support them.”

Lastly, GitHub revealed it’s rebuilding its Git architecture to decouple compute and storage, boosting throughput 35 times and enabling “virtually limitless” read throughput. “That is, again, our responsibility to the world,” Rodriguez said.

GitHub’s struggles have also opened the door for rivals eager to chip away at the company’s market share. In fact, a startup founded by former GitHub CEO Thomas Dohmke, Entire, launched this week with a platform built to store, review, and ship code written by AI agents. And it’s not the only one: Cursor launched Origin, its own code-hosting platform for AI agents, in August, while longtime alternatives such as GitLab, Gitea, and SourceHut continue to court developers. There’s undoubtedly a lot of pressure on GitHub—and the company knows it. GitHub’s answer is to invest heavily in keeping up.

GitHub’s role echoes that first flame passed from one person to the next, a tool that only matters if people keep sharing it. As AI widens the circle of who can build, Rodriguez wants developers to know where they belong. “We’re the home of developers,” he remarked. “While everyone is trying to say we don’t need developers, we’re telling them we are their home, and we’re going to be here with them forever.”


Discover more from Ken Yeung

Subscribe to get the latest posts sent to your email.