this seems like a misunderstanding of the paper: they trained a model to EDIT the external harness, not to swallow it. if anything, the paper shows harness engineering itself could be automated, to give room for the harness to adapt and unlock more of what models can do. https://x.com/Hesamation/status/21095750…

Pedro Domingos@pmddomingos16h ago

That didn’t take long: the model swallows the harness. https://arxiv.org/abs/2609.35738