Your Agent Is 80% Plumbing. Here Are the 12 Pieces You're Missing.

Anthropic just accidentally published the full source code of Claude Code, one of the most commercially successful agentic AI systems ever shipped. 1,902 files totaling 512,000+ lines across 29 subsystems. The entire architecture of a product doing an estimated $2.5 billion in annualized revenue, exposed because someone forgot to exclude a source map file from an npm package.

Every AI newsletter and Twitter thread is doing the same thing with it: cataloguing the hidden features. The Tamagotchi pet. The unreleased voice mode. The 44 feature flags. That’s interesting for about five minutes.

I mapped the infrastructure underneath the features and extracted the design primitives that determine whether an agentic system actually works in production. What I found is that the LLM call is maybe 20% of Claude Code. The other 80% is plumbing: session persistence, permission pipelines, context budget management, tool registries, security stacks, error recovery. The boring stuff that nobody writes tutorials about, and the exact stuff that separates a demo from a system that has to work when millions of people depend on it.

Every “how to build agents” tutorial on the internet stops at the demo stage: get the prompt right, wire up tool calling, ship it. And then it breaks in production in ways nobody warned you about. Sessions don’t survive crashes, tools run without permission, context windows overflow, costs spiral, and there’s no way to tell what went wrong. Thousands of developers are hitting exactly these problems right now, and the tutorials they learned from never covered them.

Here’s what’s inside:

• Two leaks, one week, zero coincidences. What Anthropic’s back-to-back exposures reveal about AI-assisted development velocity outrunning the operational discipline that’s supposed to keep it safe.

• The 12 infrastructure primitives, prioritized. Everything Claude Code runs on beneath the LLM call, organized not by how it appears in the source but by what you need to build first — day one, week one, month one.

• An 18-module security stack for a single shell command. How Anthropic thinks about permissions, crash recovery, token budgets, and session persistence at scale — and what your system should borrow.

• The confirmation nobody expected. Within hours, developers ported the entire harness to Python and Rust, proving these patterns aren’t Anthropic-specific — they’re structural requirements of any agent that has to work for real.

• A prompt and a skill to audit what you’ve actually built. An architecture audit that interviews you about your agent system and returns a gap analysis against all 12 primitives — plus a free skill package for Claude Code and OpenAI Codex that reads your codebase and tells you what’s missing.

Let me start with how this leak happened, because the meta-lesson frames everything that follows.