Inside Claude Code from its Anthropic builders
Cat Wu and Boris Cherny, the engineers who built Claude Code at Anthropic, walk host Dan Shipper through how the terminal-based agent came to be. The product was unplanned: a successor to an internal research project called Clyde, originally prototyped as a quick terminal chat while they were building on the Anthropic API. The "aha moment" was when they gave the model bash and watched it start scripting on its own; from there, the team kept adding only the tools that improved UX or enforced permissions, and recently even unshipped the LS tool in favor of a single bash + permission model. The team describes a bottoms-up, "antfooding" culture (Anthropic's term for dogfooding) in which 70–80% of technical staff use Claude Code daily, generating a feedback post roughly every five minutes. They walk through slash commands, hooks, sub-agents, and the "latent demand" design philosophy—building extensible primitives, watching how power users hack them, and then productizing the patterns. They report a growing cohort of internal users spending over $1,000/month on credits, often running large parallel sub-agent swarms for code migrations, lint rollouts, and framework upgrades. The pair also discuss memory via diary entries, plan mode, and what they call "compounding engineering"—codifying learnings from each shipped feature into prompts, sub-agents, and slash commands so the next feature is easier. On trajectory, Wu and Cherny say the CLI is not the final form factor—Anthropic is experimenting across IDE extensions, web, and mobile, and just rebranded the SDK as the "Claude Agent SDK" after seeing thousands of non-coding use cases (health, finance, legal, note-taking). They note that frontier-model autonomy is now on the order of 30 hours and heading toward days, which raises new problems around containers, monitoring, and inspecting long-running agents. Internally, Anthropic has roughly doubled in headcount since January while engineering productivity per engineer is up about 70%—but they caution that more code also means more to delete, and the team is working on a stronger culture of unshipping features that don't earn their place.
Original sources
Podcast AI & I
Episode:Inside Claude Code From the Engineers Who Built It ↗
YouTube AI & I
Video:The Secrets of Claude Code From the Engineers Who Built It ↗
AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.