Inside Anthropic's Claude Platform
Dan Shipper talks with Angela Jiang (Head of Product) and Katelyn Lesse (Head of Engineering) of Anthropic's Claude platform team about how the notion of an AI "platform" has evolved from a simple completion endpoint to stateful managed agents. They walk through the primitives behind Claude Managed Agents — the Messages API, code execution, web search, file systems, skills, and vaults — and argue that the platform is becoming a higher-order abstraction aimed at letting users specify outcomes and budgets rather than engineering harnesses. A central theme is path dependency: the choice of harness (file systems vs. reasoning chains, etc.) locks models into particular strengths, so coupling a model to its harness produces far better results than a generic hot-swappable harness. They compare approaches used by Stripe's Minions, Ramp, and Vercel's "AI software factory," and describe team-level agent patterns — like a legal-review agent watching marketing copy — that go beyond what a single Claude Code session can offer. Looking a year out, both guests describe a future where users simply specify an outcome and a budget ("Claude, make me a billion dollars, your budget is ten dollars") and Claude figures out which model, subagents, and architectures to spin up — essentially writing itself on the fly. Katelyn emphasizes that the platform team's job is to ensure the underlying infrastructure can scale to support that world of constantly running, self-recreating agents.
Original sources
Podcast AI & I
Episode:The Secrets of Claude's Platform From the Team Who Built It ↗
YouTube AI & I
Video:The Secrets of Claude's Agent Platform From the Team Who Built It ↗
AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.