AgentsAugust 7, 2026via MarkTechPost
Microsoft Open Sources code-testing-generator: a Polyglot Unit-Test Agent That Hits 92.1% Task Completion Versus 78.9% for Stock Copilot
Why it matters
A concrete agent deployment showing how agentic patterns (read → plan → execute → validate) beat prompt-and-pray on real developer workflows. The polyglot capability and internal benchmark data matter for practitioners evaluating agent frameworks.
Key signals
- Microsoft open-sourced code-testing-generator in dotnet/skills (MIT license)
- 92.1% task completion (140/152 tasks) vs 78.9% for stock Copilot (120/152) on internal benchmark
- Agent reads repo first: detects language, test framework, existing conventions, build/test commands
- Workflow: plan → write → run → validate tests
- Polyglot support (language-agnostic)
- Gain concentrated in vague prompts and diff-targeted requests
- Published August 6, 2026
The hook
92.1% vs 78.9%. Microsoft open-sourced a unit-test agent that outperforms stock Copilot on vague prompts — and it's MIT-licensed.
Microsoft has open sourced code-testing-generator, a polyglot unit-test agent shipping in the MIT-licensed dotnet/skills repository. It reads a repository before writing anything — detecting the language, test framework, existing conventions, and the real build and test commands — then plans, writes…