Anthropic's Katelyn Lesse, Angela Jiang Detail Opus 5.5 on Latent Space

The Latent Space conversation is framed around the engineering behind the Claude Platform rather than a single product announcement. Katelyn Lesse and Angela Jiang walk through how Anthropic launches models, how it approaches benchmarks and model personalization, and how it runs managed agents with memory and agent harnesses.

A recurring theme in the episode's chapter list is the split between the model's "brain" and the agent's "hands," paired with segments on protecting credentials and sandboxes and on safety classifiers and model routing. Those pieces suggest the platform's safety story is built into how requests and agent actions are routed, rather than bolted on afterward — an inference from the episode's structure, not a stated finding.

The episode also touches on infrastructure, including CPU constraints and rebuilding the Messages API in Rust, and on the boundary between training incidents and production incidents. It closes with developer relations and a segment on what to try with Opus 5.5.