The fact that he uses the term training (from the best agents), is fallacious. Rarely LLM models are fine-tuned / trained for Agents. Most of the time it is just out of the box models powering the Agentic framework, no training involved, just context, tools and maybe some other peripheral stuff like guardrails etc. That said, AI agents can be great if they have the right context for the task with limited scope presented to them, if not, they will make something up, right or wrong.