Claude Code architecture revealed: six-layer system with multi-agent support, context compression, and git worktree isolation

Claude Code is not just a CLI but a six-layer system where the Claude model is one node inside a loop. The Input Layer handles session management and permission gating; the Knowledge Layer holds skill registry, context compressor, and cross-session memory; the Execution Layer runs tool dispatch (bash, read, write, grep, glob, revert); the Integration Layer connects to external servers via MCP runtime and persists memory to markdown files; the Multi-Agent Layer supports subagents (lightweight workers in-session) and agent teams (independent Claude Code instances with git worktree isolation); and the Observability Layer wraps everything with an event bus logging all tool calls and messages.

The context compressor is a 5-layer cascade that activates when the context window hits roughly 95% capacity. Instead of summarizing conversations, it runs structured extraction on file paths, code snippets, and error histories while pruning redundant tool outputs. The prompt cache reuses stable prefixes at roughly 10% of the original cost.

Agent teams use two coordination mechanisms: a shared task list (JSON files on disk) and a mailbox system for peer-to-peer messaging. Each teammate gets git worktree isolation—a separate working directory with its own branch sharing the same repository history—so agents can write to overlapping parts of the codebase without file conflicts. Worktrees with no changes are cleaned up automatically; worktrees with changes persist for human review before merging.