Inside Claude Code's Architecture: The Agentic Loop That Codes For You

How Anthropic built a terminal AI that reads, writes, executes, asks permission, and loops until the job is done I've been living inside Claude Code for months. It writes my code, runs my tests, commits my changes, reviews my PRs. At some point I stopped thinking of it as a tool and started thinking of it as a collaborator with terminal access. So I read the architecture doc. Not the marketing page, not the changelog — the actual internal architecture of how Claude Code works under the hood. And it's more interesting than I expected, because the design decisions explain a lot of the behavior I've been experiencing as a user. Here's what's actually going on. The agentic loop Claude Code isn't a chatbot with a code plugin. It's an agentic loop. You type something. Claude responds with text, tool calls, or both. Tools execute with permission checks. Results feed back to Claude. Claude decides whether to call more tools or respond. Loop continues until Claude produces a final text response