ARTIFICIAL INTELLIGENCE
Developer Proposes 'Ephemeral Testing' to Judge Code Quality by What AI Builds on Top of It
Software performance expert Daniel Lemire outlines a new quality-assurance technique: have an AI agent build disposable applications on top of your code, then judge your work by how easily that layer comes together.
Image: Ephemeral testing · Uploaded by IntraGoals — usage rights confirmed
Software developers have long relied on a familiar toolkit to catch bugs and verify behavior: unit testing, fuzz testing, integration testing. Daniel Lemire, a computer science professor at TELUQ and a software performance expert, is proposing an addition to that list, one he says would have been impractical before the rise of capable AI coding agents. He calls it ephemeral testing.
The idea, outlined in a recent blog post, works like this. A developer writes or generates a piece of code, such as a library or software component. Rather than testing that code directly, they ask an AI agent to build something on top of it, an application or an additional layer, and have the agent test what it produced. The original code is judged indirectly, through the quality of the throwaway software built on it.
Lemire describes the technique as a variant of integration testing, with one key difference: everything built on top is disposable. Once the exercise reveals what it needs to reveal, the extra layer is discarded.
The underlying logic, he argues, is that code quality shows up in how easily other software can be built on it. A library with a clean interface, predictable behavior, and clear error messages lets an AI agent produce working code quickly. A library with hidden state, unpredictable defaults, or thin documentation tends to generate a trail of failed attempts and patches. Lemire says those failures say more about the original code than about the agent attempting to use it.
The process can be repeated with different agents and different tasks on the same foundation, giving developers a range of evidence about how their code holds up under varied use. In effect, instead of trying to anticipate every future requirement while designing a system's core, a developer can simulate those future layers by actually having them built, even temporarily.
Lemire anticipates an objection: with AI tools now able to regenerate software quickly, why not just rebuild everything as needed rather than aim for stable foundations? He counters that constant rebuilding isn't practical at scale, and that some foundational stability remains necessary. Ephemeral testing, in his view, offers a way to probe for that stability without committing to building and maintaining permanent integration tests.
Lemire says he has already begun applying the method across his own projects, using it as a quick way to prototype what a future feature might require before committing to it. He describes the results so far as promising, though the technique remains an informal practice rather than a formally studied methodology.