I particularly like Simon Willison’s definition of an agent as an LLM with the ability to call external tools in a loop.