Anthropic's platform engineers used a recent Latent Space podcast to explain how they build and ship the Claude Platform, the developer-facing layer through which companies access the company's models. Katelyn Lesse and Angela Jiang discussed the launch of Opus 5.5 and the engineering decisions behind it. A central theme was the separation of the model's 'brain' from the agent's 'hands': keeping the reasoning component distinct from the parts of a system that take action, such as calling tools or changing external systems. The pair presented this as a way to make agents more predictable and easier to control. They also covered scaling the company's APIs to meet growing demand and building managed agents, hosted services that carry out multi-step AI tasks for customers. For Australian businesses and professionals, the episode changes little immediately unless your team builds on the Claude API or managed agents. If it does, three points matter. Cost scales with usage under API pricing. Risk rises when agents can act on systems, so tool permissions should be tightly scoped and high-impact actions kept under human review. Data handling also needs checking: any personal information about Australians sent to a third-party model should be assessed against the Privacy Act 1988 first. A low-risk starting point is a pilot with read-only tool access before granting write permissions.