For Meta-Harness, I was often asked whether harness optimization actually had enough headroom. My answer so far has been "it depends" We find here that training the model between harness-search iterations can _create_ headroom.

We propose a simple recipe that resembles block ascent: alternate between (1) update _weights_ with harness fixed (2) update _harness_ with weights fixed At different points, either component can be the bottleneck, and updating it can give the other room to improve.

WHALE outperforms training (1) harness only, (2) weights only, and (3) prompt+weights, by a pretty wide margin.

We also have an adaptive variant of WHALE, where you don’t even need to tune the phase length hyperparameters. It monitors the training metric and switches when learning plateaus. This heuristic was competitive with a pretty wide hyperparameter search in our experiments.

I think this idea will scale. My bet is that large post-training runs will benefit a lot from spending even ~1% of their rollout budget on improving the harness. We should take text optimization seriously yoonholee.com/blog/2026/we-s…

Some future directions I'm excited about: - WHALE currently “co-overfits” to one good harness. How can we train models that stay robust to harness choice? - Explicitly train models to write/improve their own harness? Access to internal activations is probably useful for determining which harness will work well, especially if the weights are trained against the current best harness, as in WHALE. - Periodically distill the info in a good harness back into the weights; an agent version of arxiv.org/abs/2209.15189 @sea_snell