I Read the Claude Code Source Analysis So You Don’t Have To

A deep-dive summary of a 60-page architectural analysis of Claude Code (Anthropic's agentic coding tool), revealing that only 1.6% of its codebase is AI decision logic while 98.4% is operational infrastructure. Key findings include: the core agent loop is a simple while loop with no planning phase or state machine; a 7-layer deny-first safety architecture designed around the empirical finding that users approve 93% of permission prompts; a 5-stage context compression pipeline treating context window as the binding constraint; four extensibility mechanisms with graduated context costs; and subagent delegation using isolated contexts with summary-only returns. The post draws architectural lessons for agent builders: invest in the harness over scaffolding, treat context as the binding constraint, and design safety for inattentive users.