Microsoft Open Sourced code-testing-generatorIt is a multi-language agent who writes unit tests, then verifies that they are working. The software is delivered in the dotnet-test plugin The MIT licensees have developed a MIT software. dotnet/skills repository.
Agents fill in a hole that is often left by coding assistants. A prompt like ‘generate unit tests’ does not say which framework, file location or assertions to use. code-testing-generator It reads the repository first, before writing anything. After that, the software plans, produces, tests and runs them. Microsoft’s internal benchmark of 152 tasks was completed by 140 tasks compared to 120 with GitHub Copilot. The same prompts and model were used in both setups.
Is it deployable
Yes. Because it is an agent description and skills that are not hosted, the code remains local.
- The Company Stage: It is feasible for single maintainers and upwards. Because the agent provides repository information that small teams do not have time for, startups and midmarket teams benefit most. The language guide can be adapted to fit the internal frameworks of enterprises.
- Industries: regulated or audit-heavy software estates — financial services, healthcare, insurance, public sector — plus platform teams paying down legacy test debt.
- Applications: The backfilling of untested modules with tests, the generation of tests to compare a diff between a pull request and a change-request, increasing coverage in advance before a gate release, standardising conventions for monorepos that are polyglot.
What an agent does
It coordinates work through a Research-Plan-Implement (RPI) pipeline. It looks for tests in the code repository, it detects language and frameworks for testing, it reads tests to see if they follow conventions and then finally, the actual build and test command is found. The last step targets a failure in particular: a local test project which never gets run by CI, because it was not registered.
Choose one of the three possible strategies. You can also Direct Write and validate tests instantly. Single pass Runs one cycle. Iterative It is repeated for larger scopes of coverage or target targets. The code is never modified in production, and it avoids testing that calls external URLs or ports, or depends on timing.
Verification gate
Agents perform five tests before they can report completion. The agent uses a light form of mutation test to reason about minor code changes which should cause the tests to fail. It checks for missing or weak assertions. Every scenario is mapped to a specific test. The full workspace is built and the entire suite of tests are run. The repository confirms that the repository test command has discovered the new tests.
Benchmark results
The agent achieved a score of 95% on Microsoft’s benchmark internal test, which included 152 real-world tasks. 140 (92.1%) versus 120 (78.9%) For stock GitHub Copilot, using the same model with prompts (63%) fewer errors.
Concentrated gain. The agent scored 79 out of 89 (88.8%), compared to 59 (66.3%), reducing failures by 30 points. In 63 specific prompts both agents scored 61. The agent completed all 15 of the tasks that targeted a particular diff. Stock Copilot failed none.
The agent generates a lot of heat. Tests are 2.3% less. Comparing the two lines of coverage, there was a difference in line coverage (6.963 vs 7.129). The average task took 359 seconds versus 380. Use of tokens per completed task increased by 3.2%.
With 45.NET jobs, Claude Opus 4.8 was able to achieve 43/45 (with the agent) compared with 35/45 without; and GPT-5.5 achieved 41/45 versus 37/45. Externally, GPT-5.5 achieved 41/45 versus 36/45. Claude Opus 4.8 reached 43/45 with the agent compared to 35/45 stock. SWE Atlas Completion was measured at 16/44, compared to 12/44.
How the agent converts a prompt into a verified test
/* 4. benchmark */
SETS=[
[‘All 152 tasks’,[[‘Tasks completed’,140,120,152]The same model had a failure rate of 78.9% compared to 92.1% with this agent. This is a 63% reduction. The agent generated 2,3% less tests with the same effective coverage.’],
[‘Vague prompts’,[[‘Resolved (89 tasks)’,79,59,89]],’88.8% versus 66.3%. The number of failures dropped from 30 to 10, a 67% reduction, and the group accounted for all 20 gains.’],
[‘Detailed prompts’,[[‘Resolved (63 tasks)’,61,61,63]”96.8% of both. The workflow is not necessary when the decision-making has already been made by the developer.’],
[‘Diff-targeted’,[[‘Resolved (15 tasks)’,15,0,15][]’The agent has passed the 15 test tasks that were asked to be performed for a particular code change. Stock Copilot failed none.’],
[‘SWE Atlas’,[[‘Tasks completed (44)’,16,12,44],[‘Tests catching injected bugs’,360,316,400]],’A harder external benchmark. The completion rate is lower for both parties u2014 a 36.4% against a 27.3 %.’]
];
Bi=0
$(‘#dtagBenchTabs’).innerHTML=SETS.map(function(s,i){
Return ‘‘}).join(”);
function drawBench(){
var set=SETS[bi];
$$(‘#dtagBenchTabs .chip’).forEach(function(c){c.setAttribute(‘aria-pressed’,String(+c.dataset.b===bi))});
$(‘#dtagBench’).innerHTML=set[1].map(function(r){
Return ‘
‘+r[0]+’‘+r[1]+ vs “+r[2]+’
‘+
”+
‘
‘;
}).join(”);
$(‘#dtagBenchNote’).textContent=set[2];
setTimeout(function(){$$(‘#dtagBench .track i’).forEach(function(i){i.style.width=i.dataset.w+’%’});resize()},60);
}
$$(‘#dtagBenchTabs .chip’).forEach(function(c){c.addEventListener(‘click’,function(){bi=+c.dataset.b;drawBench()})});
window.addEventListener(‘load’,resize);setTimeout(resize,300);setTimeout(resize,900);
})();

