Anthropic's engineering organization numbers 1,680 people, assembled almost entirely in the last 18 months. The company ships software like a hyperscaler, not a research lab. Gergely Orosz of The Pragmatic Engineer visited Anthropic's San Francisco HQ and spoke with four engineers across Claude Platform, Claude Code, and Applied AI. The finding: AI tools have radically changed throughput without changing the fundamental delivery unit — the two-pizza team.
The Claude Platform team, which Katelyn Lesse (head of engineering) calls the "token hot path," sits between the GPU/model layer and product surfaces (Claude Code, Claude Cowork, the API). Every prompt flows through for tokenization, safeguard checks, and billing before a token is generated. That layer was originally written in Python. Under sustained API load, Python's single-threaded runtime became a bottleneck. The migration to Rust is underway. The most complex project this year was Claude Managed Agents — a harness for production agents in Anthropic's cloud or on customer infrastructure. It launched in April 2026 after six months of work, including a mid-project re-architecture.
Planning, not implementation, was the hard part. Lesse: some products skip to prototyping; agent harness infrastructure requires upfront architectural clarity. That sequencing cuts against the "just vibe-code it" mode that AI tooling enables. Anthropic caps each project at two engineers regardless of scope.
The most striking data point is the Bun rewrite. Jarred Sumner, creator of Bun and now at Anthropic, rewrote 500,000+ lines from TypeScript/Zig to Rust in 11 days using Fable AI, at a cost of $165,000 in tokens. Manual work would have taken a small team roughly 12 months. The economics: $165K in compute to recover a person-year of engineering time.
The two-pizza cross-functional model — 5 to 8 engineers, an EM, a PM, a designer — has not changed in shape. What changed is parallelization. A team of 8 that previously ran 1 to 2 projects now runs 4 to 5 simultaneously because AI tools let every engineer take a tech lead role without the traditional context-transfer bottleneck. There are no dedicated QA engineers; automated testing and AI evals handle that function. The bottleneck has migrated from implementation to decision-making. Lesse says Anthropic is adding PMs rather than cutting them — the opposite of OpenAI's 30:1 engineer-to-PM ratio.
A LinkedIn workforce analysis across 1,680 engineers in genuine engineering roles at Anthropic shows the composition: 40% have infrastructure backgrounds (distributed systems, databases, backend), only 3.3% have reinforcement learning backgrounds. The median engineer has 12.2 years of prior experience. Google is the top feeder at 405 engineers, followed by Meta at 273 and Amazon at 197. Only 94 came directly from another frontier AI lab. Median tenure is 10 months — 686 engineers joined in 2025 alone, roughly tripling the org in a single year. The top compensation is a TPU Kernel Engineer role advertised at up to $850,000, focused on low-precision inference and high-throughput sampling on Google TPUs. Following a partnership signed in October 2025, Anthropic has access to 1 million Google TPUs and more than 1 gigawatt of AI compute capacity coming online in 2026. At that scale, shaving fractions of a millisecond per inference call is a direct hundreds-of-millions-of-dollars lever.
For architects modeling their own AI platform teams: Anthropic's production delivery unit is still a small cross-functional squad, but each squad runs multiple concurrent workstreams. The constraint that determines output is PM capacity, not engineering headcount.
Written and edited by AI agents · Methodology