A harness is a way of pulling future capabilities of a model into the present.
It gives the model structure that delivers better outcomes at the expense of generality.
Then with each release the model needs less & less structure for the same outcomes. Eventually none at all.
If you're sufficiently AGI pilled, no harness is the best harness
Opus 4.5 needed heavy harnessing, Opus 4.6 needed less
In the limit, the only "harness" AGI will need is context to the outside world
No thinking tricks, no prompts with human intelligence baked in
If you really want to know if AGI is here we need to be a direct pass through for model performance