A Moment of Doubt, Then Magic
When Dohmke first saw GPT-3 in action, he didn’t believe it. "I thought it wouldn't work," he recalls. As a seasoned developer with a background in compilers, the idea that a model without any syntactic guarantees could produce valid Python or JavaScript seemed absurd. But then, it did work. Not just passably—shockingly well.
By summer 2020, GitHub had started experimenting internally with what would become Copilot, powered by OpenAI's Codex. Early telemetry showed that Copilot was writing 25% of the code in enabled files. "We didn’t believe the number," he says. "We sent them back to check." It turned out to be real, and growing.
The Case for Tab Completion
The now-ubiquitous tab-completion interface wasn't an obvious choice at the start. But developers were already used to IntelliSense and autocompletion in editors like VS Code and Xcode. Dohmke says the leap came from merging those habits with the power of a large language model. The goal: preserve flow state. "You have an idea, limited time, and energy. If we can keep you in the IDE, modifying and compiling, that's the magic moment."
Why Learning to Code Still Matters
With agents writing more code, do kids still need to learn programming? Dohmke says yes—100% yes. "It's like learning physics or music. You may not become a physicist or a pop star, but you need to understand how the world works."
But he adds a twist: the future of education must include how to work with AI. Not just syntax and logic, but how to prompt, validate, and collaborate with agents. "You still need the craft," he says. "But the craft is evolving."
GitHub Copilot Goes Open Source
In a major shift, GitHub is open-sourcing Copilot's VS Code integration. The decision wasn’t just ideological—it was strategic. Developers had already reverse-engineered parts of Copilot. By officially opening it, GitHub invites the community to extend it, adapt it to new editors, and even plug in different models.
"We have open source roots," Dohmke notes. "And now with Copilot, we're giving back to the same community that's made VS Code one of the most loved tools in the world."
Agents Are Here. Devs Are Still in Control.
Dohmke is careful to reframe the metaphor: AI isn’t taking the wheel—it’s more like driver assistance. And in most environments, you still need your hands on the wheel. "The role of the engineer now is to verify what the agent has done," he says. Without that understanding, the risk isn’t bad code, it's insecure code, and business consequences.
From Vibe Coding to Agentic DevOps
Vibe coding—the term for fast, expressive, AI-assisted prototyping—has exploded. But Dohmke sees two layers forming:
Vibe layer: fast, creative, expressive idea-to-prototype loops
Agentic layer: serious agents doing CI, tests, code quality, bug detection
The dream? Spend more time in the fun zone, and offload the boilerplate and bug hunting to agents.
The Future: Fragmented, Connected, Agent-First
Will it all converge into a single agent? Unlikely. Dohmke predicts a personal agent, a work agent, travel agents all linked by protocols like A2A (agent-to-agent) or embedded in tools like GitHub.
What ties it together is memory: "There needs to be governance over what belongs to the company, and what belongs to you."
Don’t Panic, Learn
Asked what he'd say to people worried about being replaced, Dohmke's answer is reassuring: "Yes, some jobs will change. But AI opens up the ability for anyone to become a developer. It democratizes access. It shortens the distance from idea to execution. And most of all, it lets us build more of the ideas we've been too busy to get to."
Key Takeaways:
GitHub Copilot started with doubt but proved to be "magic" in production.
AI code assistants extend learned behavior (like autocompletion) but now at the scale of entire functions.
Programming education still matters, but now includes how to work with AI.
Open sourcing Copilot's client is a nod to community and strategic innovation.
Software development is moving toward an agent-first architecture, with devs steering the system.
