Agent performance comes from scaffolding, not just the AI model

Original: This paper shows that agent performance depends less on prompts alone and more on the harness around them.

Source: x.com ↗