AI & I The Secrets of Claude's Platform From the Team Who Built It
499 snips
May 8, 2026 
Katelyn Lesse, Head of Engineering at Anthropic who builds Claude’s platform infrastructure, and Angela Jiang, Head of Product who designs Cloud Managed Agents, discuss platform evolution from API to stateful agents. They cover managed-agent primitives, production infrastructure challenges for always-on team agents, harness-model coupling, multi-agent orchestration, and measuring success by outcomes and budgets.
AI Snips
Chapters
Transcript
Episode notes
Anthropic Built Managed Agents After Internal Pain
- Anthropic built managed agents because their own internal teams repeatedly reimplemented cloud agent infra and hit the same problems.
- Katelyn Lesse says they decided to build it once properly so others wouldn't repeat the painful server and scale work.
Harness Engineering Is Tied To Model Performance
- The harness and model are becoming a paired unit instead of a generic hot-swappable layer, because model-specific harness engineering yields big performance gains.
- Katelyn notes evals showed drastically different performance across harness choices, so tying harness to model often pays off.
Plan For The Infrastructure Wall Early
- Expect an infrastructure wall when moving prototypes to production; plan for long-running sessions, transcript storage, sandboxing, and reliable orchestration.
- Angela warns that prototypes on a single sandbox or Mac mini often die in production unless you solve scale and resilience.