“Brendan Foody believes current base models possess sufficient reasoning capabilities for agentic work, and the primary bottleneck is learning tool use and company-specific knowledge via a small number of evaluations.”