In this unscripted live coding session, Burke and Pierce compete to build a web app celebrating VS Code contributors using unfamiliar languages and AI agents, highlighting the challenges and surprises of real-time AI-assisted development. They share practical insights on leveraging AI tools, managing agent workflows, and adapting to evolving best practices, ultimately encouraging developers to experiment and iterate with AI in their own projects.
In this live coding session, Burke, Pierce, and Olivia participate in a spontaneous coding competition where Burke and Pierce are challenged to build a web app that celebrates VS Code open source contributors. The twist is that they are not told the requirements in advance and must rely on AI agents to help them code in unfamiliar languages—Burke in Rust and Pierce in Go. The goal is to create a visually impressive site that pulls contributor data from the latest VS Code release, displays cards for each contributor, and allows users to give kudos via a heart button. The session is unscripted, with both participants figuring things out as they go, highlighting the real-world challenges and surprises of live coding with AI.
Burke and Pierce take different approaches to the challenge. Burke uses the GitHub Copilot CLI in the terminal, leveraging custom agents and skills like the front-end design skill and agent browser skill to automate as much as possible. He emphasizes the importance of planning mode and autopilot (a form of a “Ralph loop”) to ensure the agent works to completion and self-validates its output. Burke also discusses the nuances of prompting different AI models, favoring Opus 46 for orchestration and GPT-4.0/4.5 for code review, and stresses the need for simple, flat code structures that are easy for AI to manage and regenerate.
Pierce, on the other hand, takes advantage of being allowed unlimited agents and parallelization. He sets up multiple agent sessions for different tasks, such as scaffolding, deployment, and integrating features like the Copilot SDK for natural language queries about contributors. Pierce experiments with additional features like a leaderboard, theming, and even attempts to generate personalized AI videos for top contributors using the HeyGen API. He highlights the practical challenges of parallelizing tasks, especially before the project’s basic structure is in place, and shares tips on managing context and agent workflows.
Throughout the session, both developers encounter and discuss real-world obstacles, particularly around deployment. Burke’s deployment to Azure is fraught with authentication and CLI issues, illustrating that while AI can handle much of the coding, deploying to production environments remains a significant hurdle. Both participants reflect on how AI agents prefer monolithic, easily replaceable code over traditional abstractions, and how developers need to adapt their workflows and expectations when collaborating with AI.
The session concludes with reflections from all three participants. They emphasize that developers should experiment with agents, start small, and gradually add complexity as they become more comfortable. The rapid pace of AI advancement means that workflows and best practices are constantly evolving, and what didn’t work a few months ago may work now. The key takeaways are to give agents a real chance, leverage planning and validation features, and not be intimidated by the growing list of AI tools and features. The event wraps up with a sense of excitement about the future of AI-assisted development and encouragement for others to explore and iterate with these new tools.