When GitHub’s Copilot Coding Agent (CCA) first became available in May 2025, the promise of cloud-based AI coding agents was generating enormous excitement, and equal skepticism. So we began an experiment, and like many experiments in software engineering, it started with a simple question: could a cloud-based AI coding agent meaningfully contribute to one of the most complex, most scrutinized, and most critical open source codebases in the world

Microsoft DevBlogs

If there’s one lesson from this experience, it’s this: preparation matters more than the model.

Our success rate jumped from 38% before our instruction file to 69% after. Better preparation was the primary driver. Every hour spent documenting “how we work” in .github/copilot-instructions.md and teaching new skills pays dividends across every future CCA PR. Instructions aren’t just documentation; they’re leverage. They’re the difference between a junior developer who needs constant hand-holding and one who can work semi-autonomously.