Microsoft's .NET team built an open-source polyglot unit-test agent called code-testing-generator, available in the dotnet/skills repository. Unlike a plain Copilot prompt, the agent first learns the repository's language, test framework, conventions, and build commands before writing tests. It selects one of three execution paths (direct, single-pass, or iterative) based on task scope, then validates tests with lightweight mutation testing, assertion checks, and a full CI-compatible test run. Benchmarked against 152 internal tasks, the specialized agent achieved a 92.1% completion rate versus 78.9% for stock GitHub Copilot, with 63% fewer failures — gains concentrated on vague prompts. It supports .NET, Python, Go, TypeScript, Java, Rust, and more, and is usable today via GitHub Copilot CLI and VS Code.
Table of contents
What happens after the prompt Copy linkWhat we measured Copy linkWhat we learned and what comes next Copy linkTry the agent Copy link22.3K Impressions2 Comments