Claude Code's source code leaked, exposing the full architecture of its agentic loop: context loading, tool execution with a layered permissions model, sub-agent spawning, and multi-agent workflow orchestration. This is not speculation or reverse engineering from API responses. This is the actual implementation.
The more interesting material starts at 12:32. Unreleased features include 'Buddy', an apparent persistent companion mode, 'Away Summary' for asynchronous task reporting, server-side tools not yet exposed through the public API, and 'Auto Dream', a memory consolidation feature that likely runs background summarization to manage long-horizon context. These are not announced roadmap items. They exist in the code right now.
The video is worth watching in full for the tool permission model breakdown at 04:36 and the multi-agent workflow deep dive at 10:54. Understanding how Claude Code gates tool access and coordinates sub-agents tells you more about Anthropic's production AI architecture than any blog post they have published. Auto Dream alone raises serious questions about how context and memory will be handled in the next generation of coding agents.
[WATCH ON YOUTUBE →]