Agent Anatomy

On March 31, 2026, Anthropic accidentally shipped a source map in their npm package. 513,000 lines of Claude Code's TypeScript — exposed. I spent the next few days reading through every layer, cross-referencing with Berkeley's CS294 Agentic AI course, claw-code (Rust port), Anthropic's official docs, and the Agent SDK. **What's in the ZIP (20 files, ~3,300 lines of Markdown):** - 12 study notes covering 4 agent layers: tool loop, memory, context compression, coordinator - 5-source cross-validation audit with 64 file:line citations - 5 exercises that force you to trace real code with concrete values - Blind A/B experiment: structured vs free-form source reading (6-0 result) - 9 verified claims with full evidence chains **The 4 key discrepancies (the gold):** 1. Compaction: claw-code = deterministic; CC TypeScript = LLM with 9-section prompt 2. Agent spawning: SDK = dataclass; CC = "never delegate understanding" + cache forking 3. Permissions: docs = 3-tier auto-mode; CC = fail-secure + deny reason as is_error 4. Memory: SDK = a flag; CC = 2-layer architecture + autoDream + semantic search Pairs with the open-source repo: github.com/zl190/nano-agent-anatomy

Found on Devadex — the discovery index for independent software the big search engines bury. More from gumroad.

Report this listing