Anthropic separates Claude's brain from its hands in agent architecture
AILatent Space interviews Anthropic's heads of Engineering and Product about how Claude Opus 5.5 was launched and built. The discussion covers Managed Agents, which keep the session log and harness apart from the sandbox that executes code, so a failed sandbox does not end the agent session and credentials can stay out of generated code's reach.
AIWhy it matters.
The episode explains how Anthropic separates the agent's brain from its execution sandbox, a design choice that bears on reliability, security, and agent runtime planning.
Loading source posts…
Source details
Latent Space · swyx .ai · latent.spacePublished · added here
Original title: How Anthropic ships Claude with Claude: Separating the Brain from the Hands