I like to think of an agent as a LLM chat loop with tool calling. The agentic behavior emerges with these ingredients.

Don’t really a harness…unless you consider a shell script that runs that loop a harness.