Inside Anthropic's engineering: the AI lab keeps dogfooding, quietly

In plain terms - The Pragmatic Engineer's deep dive on Anthropic engineering says the AI lab is doing "ever more" code review and testing with AI, while two-pizza teams remain very much alive. Fewer surprises than the hype implies - and that's the interesting part.

Context

The Pragmatic Engineer publishes on July 28, 2026 a « deepdive » on Anthropic's internal engineering practices. The interest is twofold: Anthropic is the provider of the most used AI model on the dev-tooling side (native Claude Code, Cursor via API, Windsurf, etc.), and it is one of the few reliable sources on « what a company that ships its product by dogfooding its own LLM really looks like ».

The data

The post (newsletter.pragmaticengineer.com) reports the main points, summarized by its author:

• More and more code review and testing by AI.
• Two-pizza teams very much alive - the classic agile structure holds.
• Organization and process details beyond the « pure research lab » caricature.

Under the hood

The phrase « ever more » (« more and more ») is more interesting than it seems: it indicates a gradual and measured adoption, not a radical shift. Anthropic - which builds the model - has not replaced its engineers with agents, nor eliminated human code reviews: both layers coexist, and AI is gradually gaining ground.

Analysis

Three takeaways. Cadence: internal LLM adoption is a smoothing, not a big bang - consistent with Google (Amplified Engineer), Shopify (AI mandate). Structure: two-pizza teams hold because they are an organizational formality, not technical - AI changes the dev output, not the optimal size of a team. Auto-dogfood: Anthropic sells Claude Code to professional engineers; the fact that Anthropic uses it (and improves it) internally is an implicit moat against OpenAI, whose Codex stack was later grafted onto ChatGPT (see openai-super-app thread).

Scenarios (12 months)

• 60% - Other AI labs (OpenAI, Google DeepMind) in turn publish their internal practices, pushed by dev-tooling transparency pressure.
• 25% - Anthropic's internal metrics (review time, post-merge bug rate) eventually leak.
• 15% - Partial rollback: reintroduction of human safeguards on certain code classes (crypto, security, distributed systems).

Implications for the practitioner

For a lead engineer: the reading provides concrete references to justify LLM adoption to management - Anthropic has not broken its teams. For an architect: do not redefine your structure to « accommodate » AI; let it slip into existing roles. For a CTO: the real differentiator is not « which AI » but « at what level of the loop » (review, testing, deployment).

To watch

The continuation of the TPE series. The release of a dedicated Anthropic Engineering Blog. A first public post-mortem of an incident where AI might have let a mistake slip through.