Important discussion. Measuring models against harnesses is completely broken. I prefer to test model quality against minimal harnesses like Pi and Hermes Agent. This is not perfect, as there are biases in the harnesses that favor some models and not others.

A standard way to do this is missing but important, as harness engineering is where leading AI companies are focusing efforts. Not enough effort here, as things are moving fast and harness optimization is individualistic.

On the flip side, I feel like models will eventually have the ability to dynamically generate harnesses on the fly as per task. Claude models do this already to some extent, though pretty inconsistently and remain a mystery. But this could mean that a harness is just a tunable artifact like a system prompt. In that realm, how are we assessing it, and exactly what? Benchmarking will only get murkier from here onwards.