Anthropic's 200-person platform team: long-running agents make infrastructure the differentiator

Key points

Key takeaways from the Builders (Grit, Kleiner Perkins) interview with Anthropic platform leads Angela Jiang (Head of Product) and Katelyn Lesse, Head of Engineering (August 2026):

The shift. A year ago Anthropic products were human-in-the-loop chatbot sessions; today models work autonomously for long stretches, so the constraint has moved to the infrastructure that keeps agents running, recovers them from errors and keeps their work secure and compliant.

Undifferentiated infrastructure. Anthropic's managed-agent philosophy tells builders to focus on what differentiates them — customer insight, problem space, high-level tuning — and leave the hard distributed-systems work of running agents at scale to the platform.

Architecture. Anthropic separates the agent's durable harness — the "brain" — on a persistent server from ephemeral execution sandboxes: a sandbox can die or be torn down without killing the whole agent.

What changed for work. Tasks moved from narrow instructions ("edit these spreadsheet cells") to outcome-level briefs ("build me a DCF of this company and say whether to invest"): the agent opens the file, calculates, checks its own work and retries — no human babysitting.

On-ramp for enterprises. For Fortune 500 executives, Claude managed agents are the single easiest starting point — higher-order abstractions with access to lower-level primitives (prompt caching, context engineering) as teams grow more opinionated.

Tokens as advisors. Katelyn on strategies that work: "Use tokens for different jobs, not just executing, also advising."

The team. Both joined within the past year — Angela via Stripe's v2 API and OpenAI, Katelyn from Stripe Connect — attracted by talent density and a team "really ready to get punched in the face" on safety, infrastructure and scale problems.

Read more: Builders S2E1 — Grit (YouTube) · Kleiner Perkins