Claude Code's Architecture, explained visually!

Claude Code is a six-layer agent harness, not just a CLI wrapper around Claude. The layers cover: Input (session management, permission gating, YAML trust tiers), Knowledge (skill registry, context compressor, task graph, cross-session memory), Execution (typed tool registry, streaming runtime, prompt cache), Integration (MCP runtime connecting to external servers, persistent agent_memory.md), Multi-Agent (subagents for lightweight parallel tasks vs. full agent teams with git worktree isolation and a mailbox messaging system), and Observability (event bus with lifecycle hooks for full audit trails). At the center is a deliberately simple 'dumb loop' — single-threaded, assembling context, calling the model, executing tool requests, and repeating — while all intelligence lives in the surrounding layers.