We gave two small LLMs a real customer-service task. One got 100/100. The other stopped with the job half-done. I wanted to test something more meaningful than benchmark scores: can a small model actually complete a multi-step agent task without handing the work back to the human? The task: A custom
