Daniel Lemire proposes 'ephemeral testing': judge your library by building throwaway layers with AI

lemire · x · 2026-10-05

Computer scientist Daniel Lemire proposes ephemeral testing, a software quality method that was unthinkable before AI agents:

Write your component, then have an AI agent build applications or extra layers on top of it and test what it built. You assess not the original work directly, but the quality of software built on top.
It's a form of integration testing where the upper layers are entirely throwaway.
The signal: a library with a clean API, stable invariants, and useful errors lets agents produce working software quickly; hidden state, surprising defaults, or incomplete docs produce piles of patches and failures — and those failures are evidence about your code, not the agent.
Repeatable across different agents and tasks on the same foundation.
Instead of anticipating upper-layer needs while building the core, you simulate them by actually building them.

Lemire says he applies this trick across projects, prototyping what might later be built atop a proposed feature before committing to it.

More from coding & agent