Summary
Jess Yan, product lead at Anthropic for Claude Managed Agents, sits down with Peter Yang to explain how Anthropic’s approach to agent infrastructure has evolved and what the company is building for developers who want to deploy autonomous AI systems. The conversation traces the arc from simple prompting loops to long-running, self-discovering agents with access to third-party systems, sensitive internal data, and the need for robust permissioning, observability, and steering.
Yan introduces Claude Managed Agents as a pre-built harness and companion infrastructure designed to dramatically reduce the effort required to deploy production-grade agents. Key design goals include easy-to-stack developer primitives, out-of-the-box infrastructure for tasks that run overnight or across extended time horizons, and built-in support for human-in-the-loop checkpoints. Anthropic uses these systems internally: Yan describes a predictive customer model that generates rich analytical insights in minutes, and overnight agent workflows where bugs are autonomously resolved and backlogs cleared while engineers sleep.
The interview also explores why model development and harness design can’t be fully decoupled — Anthropic tests models in conjunction with specific harnesses, which keeps the two tightly paired — and how the nature of agent outputs is shifting. Rigid structured JSON outputs are giving way to outcome-optimized agents that self-correct along the way, with the end goal (a beautiful interactive artifact, a specific accuracy benchmark) replacing step-by-step output specifications.
📺 Source: Peter Yang · Published June 28, 2026
🏷️ Format: Interview
