How Does Claude Code Actually Work? Let's Make Our Own

The video begins by introducing Claude Code as a popular AI tool that many people don't understand how it works. The presenter aims to demystify it by explaining its core components and the underlying architecture. The agent's operation is described as a loop: building context, model prediction, guardrail checks, and tool execution, with results fed back into the context.

The five building blocks are detailed: 1. **Model:** The LLM itself, which simply predicts the next token and remembers nothing. It only sees the context window. 2. **Tools:** These are how the agent interacts with the world, categorized into built-in (IO tools like read/write files, bash) and external (MCP servers, plugins). The key idea is that the model asks for tools, but the software runs them. 3. **Memory:** This is effectively just files, specifically markdown files in a `/.claude/memory` directory, which the agent reads and stuffs into the context window when needed. 4. **Context Window:** This refers to all the information the model can currently see, essentially a very long prompt that includes the system prompt, tool definitions, skill definitions, memory, conversation history, and tool calls/results. 5. **Guardrails:** These are software-enforced rules, not part of the model, that decide what actually runs. This includes human-in-the-loop approval for risky actions and limits on what the agent can do.

The video then delves into the implementation details of each component, showing Python code for the agent's loop, tool discovery, and how the model interacts with tools by outputting structured text (JSON) that the harness then parses and executes. Finally, the presenter shows a simplified Python code example of the agent's loop, highlighting its simplicity and the role of each component.

The video concludes by summarizing that it's all just text moving through a loop, and every feature seen—skills, memory, approvals—is one of these five blocks wearing a different hat. The model simply predicts text, and the harness handles everything else.