New method predicts AI agent performance cheaply from simple tasks

Original: Agentic evaluations are sloooow, and expeeeensive.

Source: x.com ↗