Everyone is using SOTA models for software, but for a commercial application that requires cost-effectiveness, if you want to fine-tune a weak or mediocre model into an agent that can perfectly complete tasks, you'll need a lot of repetitive benchmarking work.