Claude Code Leak Reveals: It's Not an API Wrapper, but a Full Operating System

Claude Code leak revealed: it's not a wrapper over API, but a full-fledged operating system

The leak of Claude Code's source code — Anthropic's console AI agent — revealed much more than commentators discussing it in the first days expected. 512,000 lines of code and 44 feature flags were published, and along with them — a pet game system in the spirit of Tamagotchi with internal names like Tengu, Fennec and Penguin mode. As the author of the material on Habr writes, while the internet argued whether this was a toy or serious architecture, something far more important turned out to be the case: the leak for the first time showed how AI agents that bring Anthropic 2.5 billion dollars in annual revenue actually think.

What exactly leaked

The leak itself is not just a couple of files, but a voluminous snapshot of the internal structure of the product: over half a million lines of code and dozens of functionality switches, each of which enables or disables a separate capability of the agent. The presence of 44 feature flags in itself speaks to the maturity of engineering culture inside Anthropic — the team clearly manages the rollout of changes gradually, testing new mechanics on a portion of users, rather than rolling everything out at once.

Game elements — Tamagotchi-like pet system with names Tengu, Fennec and Penguin mode — are what first spread across social networks and other publications: entertaining details are easy to quote and discuss. But, in the opinion of the Habr author, focusing on them distracted from much more significant engineering infrastructure hidden behind cute names.

What is known about the leak:

• Volume of leaked code — 512,000 lines
• Number of feature flags — 44
• Game layer — pet system in Tamagotchi style with names Tengu, Fennec, Penguin mode
• Estimated annual revenue of the product — 2.5 billion dollars
• Key conclusion of the author — Claude Code is not organized as a wrapper over API, but as a full-fledged operating system for an AI agent

Half a million lines of code is a volume comparable to large independent software products, not a thin layer that simply forwards user requests to a language model API and returns a response. It is the scale of the leak that prompted the Habr author to reconsider the widespread belief before that console AI agents are primarily a convenient interface over someone else's model, rather than an independent and complex engineering system.

Three engineering patterns behind the game facade

Behind the names of pets, according to the author's description, lies a hard engineering reality — three patterns that transform Claude Code from "an expensive autocomplete" into a serious product. The first is a self-recovering request cycle: a mechanism that allows an agent to not stop at a single step failure, but to correctly resume work and bring the task to completion, even if something goes wrong at an intermediate stage. The second is dream computation, that is, executing part of the processing not at the moment of direct user interaction, but in a background, "deferred" mode, which allows the product to work more efficiently without constant load on live dialogue.

The third is a two-level function cutoff system that determines which agent capabilities are available at a given moment depending on context, task execution stage, or flag system settings.

Why this changes the conversation about AI agent architecture

The main thesis of the material is that what at first glance appears to be an amusing detail of a product, a Tamagotchi, funny mode names, in reality is the visible part of a serious engineering system for managing complex agent behavior. Anthropic accidentally exposed not a list of marketing features, but the working architecture of one of the most successful agentic products on the market — with annual revenue of 2.5 billion dollars, which puts Claude Code among the most commercially successful AI developer tools.

For engineers building their own agents, the leak is a rare opportunity to see how a real industrial system solves the problems of reliability (what to do when a request fails), efficiency (how to distribute computations over time) and complexity management (how to safely enable and disable functionality for different users). It is these three patterns, not game mechanics, in the author's opinion, that explain why Claude Code managed to grow from a code autocomplete tool into a full-fledged operating system for an AI agent.