What actually makes Claude Code smart? Not just the model, the harness around it. To prove it, I built my own harness in pure Python from scratch. Start with a raw LLM call and it's useless: no tools, no memory, no clue about your files. Add an agentic loop (about ten lines of Python) and it can chain tool calls, list and read your files, and tell you what to do.
Add an MCP client, the same protocol Claude Code and other harnesses use, and it instantly picks up tools you never wrote, like live web fetching. Add persistent memory through context engineering, and you can even swap the model mid-conversation to a free local one and it still remembers everything.
#techwithtim #ai #aiengineer