Microsoft's Test Agent Hits 92.1%, Beats Copilot

Why this matters

The 92.1% versus 78.9% gap matters less than how Microsoft closed it. Stock Copilot fails here because "write unit tests" is an underspecified prompt, and the model guesses at framework, file placement, and assertion style. code-testing-generator instead reads the repo first, then runs five verification checks including mutation-style reasoning about which code changes should break its own tests. That's a meaningful design shift: verification baked into the agent's workflow rather than left to a human reviewer after the fact.

For developers, the MIT license and the dotnet/skills packaging mean this is inspectable and forkable today, not a locked-away benchmark claim. Founders building dev tools should note where the real gains came from: not a bigger model, but a scaffolding of self-checks around a modest one. That's a cheaper, more portable lesson than "use a better LLM."

For researchers, the open question is whether this verification pattern generalizes past unit tests, to code review, refactoring, or documentation agents, or whether test generation just happens to have unusually clean success criteria that other coding tasks lack. Worth watching who forks this next and what they build on top of it.