We deployed a coding agent that hit 94% on the industry benchmark. It failed in production on the first real edge case because the benchmark measured single-turn success and our actual work was multi-turn refinement. The model could not update its beliefs correctly when new evidence arrived, something no single-turn eval would catch.
This is not a hypothetical. I have watched agents shine in demo and disintegrate on the messy input that production actually serves. The gap between what we measure and what ships is real, and it is where reliability lives or dies.
The benchmark misses the point
FutureBench evaluates agents by asking them to predict events that occurred after their training cutoff. This removes the possibility of correct answers coming from memorized training data rather than genuine reasoning. The design matters because it tests whether an agent can reason, not whether it can recall.
BayesBench showed that standard LLM evaluations score only final-turn answers in single-turn format, leaving multi-turn belief updating entirely unexamined. Across seven models, scaling improves latent inference and evidence accumulation but LLMs do not match rational Bayesian updating. In production, your agent runs many turns. The benchmark that stops at turn one is not measuring the thing that actually breaks.
