Why is there no serious resource on building an AI agent from scratch?

Not wrap the OpenAI API and slap LangChain on it tutorials. I mean actually engineering the internals like the agent loop, tool calling, memory, planning, context management across large codebases, multi-agent coordination. The real stuff. Every search returns the same surface level content. Use CrewAI. Use AutoGen, cool but what's actually happening under the hood and how do I build that myself from zero? Solid engineering background, not a beginner.