이미지: METAL LAB 생성
Summary
- Microsoft has released its polyglot unit-testing agent "code-testing-generator" under an MIT license in the dotnet/skills repository
- On an internal benchmark of 152 tasks, it achieved a 92.1% completion rate (140 tasks), ahead of stock GitHub Copilot's 78.9% (120 tasks)
- The gap was widest on 89 ambiguous prompts, where it scored 88.8% versus 66.3%, and it swept all 15 diff-specific tasks 15-0
- 공개 형태
- MIT 라이선스, dotnet/skills 저장소 내 dotnet-test 플러그인
- 내부 벤치마크 완료율
- 92.1%(140/152) vs 스톡 Copilot 78.9%(120/152)
- 모호한 프롬프트(89개)
- 88.8%(79개) vs 66.3%(59개)
- diff 타깃 과제(15개)
- 에이전트 15/15, 스톡 Copilot 0/15
- .NET 45개 과제(Claude Opus 4.8)
- 43/45 vs 스톡 35/45
- 테스트 수·커버리지·소요시간
- 테스트 2.3% 감소(6,963 vs 7,129), 커버리지 동일(72.4% vs 72.2%), 359초 vs 380초
Microsoft has open-sourced "code-testing-generator," a unit-test generation agent that supports multiple programming languages. It is distributed under an MIT license as a dotnet-test plugin within the dotnet/skills repository, and rather than running as a separate hosted service, it operates inside an existing coding agent, with code reportedly staying local.
A repository-first approach
For ambiguous prompts like "generate unit tests," the agent decides matters such as framework choice and file placement by first reading the repository itself. It detects the language and test framework, identifies existing testing conventions, and checks the actual build and test commands before writing any tests. It follows a Research-Plan-Implement (RPI) pipeline covering planning, writing, execution, and verification, and Microsoft says it does not modify production code.
Benchmark results
On Microsoft's internal benchmark of 152 tasks, the agent completed 140 (92.1%), ahead of stock GitHub Copilot using the same model and prompts, which completed 120 (78.9%). The gap was most pronounced on ambiguous prompts: the agent handled 79 of 89 (88.8%), cutting failures from 30 to 10 compared with stock Copilot's 59 (66.3%). On 63 detailed prompts, however, both matched at 61 (96.8%).
On 15 tasks targeting specific pull request diffs, the agent passed all of them while stock Copilot passed none. On 45 .NET tasks, the agent scored 43/45 with Claude Opus 4.8 versus stock's 35/45, and 41/45 versus 36/45 with GPT-5.5. On the external SWE Atlas benchmark, it scored 16/44 versus 12/44.
According to Microsoft, the agent generated 2.3% fewer tests overall (6,963 versus 7,129) while achieving essentially the same coverage (72.4% versus 72.2%), with an average task time of 359 seconds, slightly faster than stock.
