Progress in AI has depended on the infrastructure that prepares data and supports training. Agent harnesses bring a similar systems challenge into execution: assembling context, coordinating tools, and verifying outcomes. How much can these choices improve performance with the model held fixed? In the attached essay, I explore this parallel, examine recent results, and break down what can be optimized across the harness. I also consider where these changes could help open-weight models to close the gap with proprietary frontier models on specific tasks. Feedback welcome!

Thanks Bikash and Vivek for perusing the early versions of the essay.


Download Paper