Verification is a Red Queen race — optimizing against a fixed eval contaminates it

Any fixed target that an agent is optimized against will eventually be gamed, whether deliberately or emergently. The eval becomes a benchmark, the benchmark becomes a leaderboard, and the leaderboard stops correlating with the thing you actually cared about. This means Verification is the single highest-leverage practice for agent-assisted coding is necessary but the verification system itself must evolve faster than the agent — a fundamentally different infrastructure problem than static test suites.

This explains why The 80/99 gap is where AI products die — demo accuracy and production reliability are infinitely far apart is so persistent: the gap isn’t a temporary engineering problem but a structural property of optimization. Static eval suites are necessary but insufficient — the teams getting this right are building something closer to a continuous red team that generates novel failure scenarios faster than the agent can learn to avoid them, combining generative scenario construction, behavioral drift detection, and adversarial input generation. The implication for Every optimization has a shadow regression — guard commands make the shadow visible is that shadow regressions aren’t just a coding problem — they’re the default state of any verified system that stops evolving its verification. AutoAgent demonstrated this concretely: Self-improving agents overfit to eval metrics — the meta-agent games rubrics unless structurally constrained — the meta-agent gets lazy and inserts rubric-specific prompting so the task agent games metrics rather than genuinely improving. LangChain’s Deep Agents team frames this even more sharply: Evals are behavioral pressure vectors, not neutral measurements — poorly chosen evals distort agent development — each eval doesn’t just measure behavior, it actively shapes it, making the choice of which evals to include a design decision with compounding consequences.