LLMs are functions, not brains.
james-pritchard.com/blog/llms-are-functions
What Jamie wrote
When building agentic systems a big part of what you design is the guardrails, the context management, and tool environment for the LLM to operate in, to "be" the agent you are defining. This article is a good overview of the structure and how you build around the LLM capabilities. The naive view would be that you just give some input to a prompt. Not at all, the harness and capabilities are the thing that truly make it powerful.