Source time Edition date
Close readLatent Space
How Anthropic ships the #1 LLM in the World safely: Separating the Brain from the Hands
Anthropic 如何安全交付全球第一的 LLM：分离大脑与双手
Anthropic's split between Claude Opus 5.5's reasoning and its agent tooling raises a real tradeoff: safety guarantees versus how much autonomy managed agents can be trusted with.
Judgment
Platform and security teams weighing Claude Managed Agents now decide how much credential scope and network reach to hand the harness, because Anthropic's split of brain from hands pushes the containment burden onto their own sandbox and trust-grant configuration rather than the model.
Basis
On Latent Space (107 min), Katelyn Lesse and Angela Jiang explain the Claude Platform split behind the Opus 5.5 launch: the model stays the "brain" while the harness and sandbox act as the "hands," with isolation, credentials and network controls as the containment layer. The mechanism in the show notes is decoupling brain, harness and sandbox, so a managed agent's tool calls run under scoped grants instead of model-level trust; the episode also covers rebuilding the Messages API in Rust (56:05) and safety classifiers plus model routing (1:12:25) as production levers, with trust grants at 26:25 and model launch gating at 30:25. The stated payoff is reliable long-running agents; the cost lands on whoever has to define those grants and sandbox boundaries before shipping.
- Model (brain)tool calls
- Harnessexecutes
- Sandboxscoped access
- Credentials
Unclear The source gives no containment numbers: no sandbox-escape or incident rates, no latency or cost penalty for isolation, and no Opus 5.5 pricing figures despite flagging pricing at 7:30.
Next Read Anthropic's "How Anthropic contains Claude" and "Scaling Managed Agents" notes and check whether credential scoping is enforced per-tool or only per-session.