One signal

Update

What changed

Anthropic executives Katelyn Lesse and Angela Jiang say the Opus 5.5 platform work balances model intelligence against cost. That puts the launch in practical terms: the question is not only what the model can do, but how much capability users get for the resources it takes.

They also describe changes to API integrations intended to mitigate prompt distillation and protect reasoning blocks. The discussion gives a more specific account of the platform work than a general promise of safer agents, although it does not quantify how well those measures work.

Lesse and Jiang describe separating the agent’s “brain” from the sandbox execution “hands”, with security and latency as the aims. The architecture detail adds a useful distinction: the model and the environment in which an agent acts are separate parts of the system.

Sources and evidence

• How Anthropic ships Claude with Claude: Separating the Brain from the Hands: Anthropic executives Katelyn Lesse and Angela Jiang discussed Opus 5.5 platform architecture, including intelligence-per-dollar trade-offs, API integration changes intended to mitigate prompt distillation and protect reasoning blocks, and separating the agent model from sandbox execution to improve security and latency.

Independent WittyWires Watcher; not an official account or feed.