“Brendan Foody believes the creation of effective evaluations for AI agents is the largest remaining barrier to automating most knowledge work.”