Latent Space: The Agent Harness Is Being Absorbed Into the Weights, and What's Left Will Be a Harness for Human Attention
Dan McAteer's August 22 Latent Space essay argues the Christmas 2025 jump in agent usefulness was not a model event or a harness event but the two improvement curves crossing. He traces four stages: ReAct as the loop on paper (Oct 2022), AutoGPT and BabyAGI as premature autonomy where the harness ran ahead of model capability, Cursor and Copilot pulling the harness back below the model curve with a human in the loop, and the current phase where models absorb harness scaffolding into post-training and engineers delete what got absorbed. The concrete number worth keeping: 95% per-step reliability across a 20-step task compounds to roughly 36% success, which is why a loop below a capability threshold amplifies errors instead of results. He quotes Lukasz Kaiser saying the winter change was hard to attribute to any single component.
Source
↳ Follow the thread