Blog

A benchmark for whether an agent gets better with experience, by remembering the puzzles it has met or by getting better at new ones. On an open model, the notes its harness kept did neither, and training its weights did both.

Harnesses are how today’s agents remember anything at all. They are a point-in-time fix, and continual parametric learning will replace them.

We built our agent, unify-agent, around what the models of 2025 couldn't do. Stronger models have since made much of it unnecessary, and we think the next step is learning in the weights.