Anthropic is separating what it calls an agent’s “brain” from its “hands” in the infrastructure behind managed agents. In an interview with Latent Space, engineering and product leaders described keeping the session log and orchestration harness outside the sandbox that executes generated code.
That boundary changes both reliability and security. An execution environment can be created only when needed, and a failed or discarded sandbox does not have to erase the wider agent session. Credentials can remain outside the code-running environment and be attached through a controlled proxy, leaving generated code with no secret to read directly. Network controls can further restrict where the sandbox connects.
The team said an earlier design placed the whole coding agent behind a sandbox API. Starting both the environment and agent increased time before the first response, while losing the sandbox also meant losing the running agent. The newer architecture lets Anthropic work on lower latency and session durability without removing execution isolation. For other agent builders, the concrete design question is which state must survive and which operations should occur in a short-lived, untrusted environment.