Source time Edition date
IndustryLatent Space
How Anthropic ships Claude with Claude: Separating the Brain from the Hands
Anthropic 如何用 Claude 交付 Claude：将大脑与双手分离
Anthropic's Claude Platform splits the 'brain' from the 'hands' on Opus 5.5, so readers should judge whether that separation buys reliability or just adds coordination overhead engineers must own.
Judgment
Platform engineers choosing an agent runtime must decide whether to split the persistent session/harness from sandbox execution, accepting added coordination and a credential proxy to survive sandbox failures and cut first-response latency.
Basis
Katelyn Lesse and Angela Jiang, Anthropic's heads of Engineering and Product, told Latent Space that Managed Agents separates the session log and harness from the sandbox that executes code. The harness can persist while an execution environment is created when needed, so a failed sandbox need not take the whole agent session with it. They also describe credentials held separately and attached through a proxy, with controls on where sandbox traffic can go, so generated code runs without a service credential in the sandbox. The guests say an earlier approach put the whole coding agent in a sandbox behind an API; starting the environment and agent added work before the first response, and if the sandbox disappeared, so did the running agent.
Unclear The source does not quantify the added coordination cost, compare latency or reliability with the earlier sandboxed design, or say which team owns the credential proxy.
Next Check Anthropic's Managed Agents documentation for whether the credential proxy and sandbox network controls are generally available and what platform engineers must configure.