Microsoft Open-Sources code-testing-generator, a Unit-Test Agent Achieving 92.1% Task Completion

Microsoft has open-sourced code-testing-generator, a polyglot agent that writes unit tests and validates they work. Available as a dotnet-test plugin in the MIT-licensed dotnet/skills repository, the agent addresses a gap coding assistants typically leave open—decisions about which framework to use, file location, and assertion strategies. The agent reads the repository before writing anything, then plans, writes, executes, and validates the tests it produces. On an internal benchmark of 152 tasks, it completed 140 tasks versus 120 for stock GitHub Copilot using the same model. The agent coordinates work through a Research-Plan-Implement (RPI) pipeline that detects language and test framework, reads existing tests for conventions, and locates actual build and test commands. It implements three strategies: Direct writes tests immediately, Single pass executes one cycle, and Iterative repeats for larger scopes or coverage targets. Before reporting completion, it runs five validation checks including lightweight mutation testing and assertion verification. Gains are largest with vague prompts—88.8% versus 66.3%—compared to detailed prompts where both reach 96.8%. Notably, the agent generated 2.3% fewer tests while maintaining identical line coverage, averaging 359 seconds per task.