Anthropic engineers explain how Claude managed agents and Opus 5.5 were built

Katelyn Lesse and Angela Jiang of Anthropic appeared on the Latent Space podcast to discuss the engineering behind the Claude Platform. The episode covers the launch of Opus 5.5, how the company scales its APIs, and how it builds managed agents that behave reliably.

A central idea is splitting the model's "brain" from the agent's "hands." In practice, that means decoupling the model, the harness that runs it and the sandbox where actions happen. The discussion also touches on protecting credentials, isolating sandboxes, and how agents keep progress across long-running tasks and multiple context windows. Related topics include memory, evaluating agents, safety classifiers and model routing, and rebuilding the Messages API in Rust.

For marketers and content teams, the interview is a look at the infrastructure that sits under agent tools they may already use or evaluate. Questions about where credentials live, how an agent is contained and how its work is evaluated matter when an agent touches a CMS, an ad account or customer data. The episode also addresses fighting AI slop, a concern for any team publishing at scale with model help.

The episode opens with a newly recorded introduction on Anthropic's S-1, covering revenue per employee and the improvement and investment loops. Teams planning agent workflows on Claude may want to review the linked engineering posts on scaling managed agents and long-running harnesses, and to try Opus 5.5 on a real workflow, a topic the episode closes on.

Sources

This brief is our own summary. Read the full reporting at the source.

• Latent Space: How Anthropic ships the #1 LLM in the World safely: Separating the Brain from the Hands