Next, you can now control where inference runs for Claude Managed Agents. Set mo

Anthropic's new `inference_geo` setting for Claude Managed Agents controls where an agent's model computation runs, not where its data is kept. Managed Agents is Anthropic's hosted harness for running Claude as a long-running agent with its own sandbox, tools and session history. The August 7 announcement offers two values: `global`, which runs wherever there is capacity at the standard rate, and `us`, which keeps inference in the US and bills at 1.1 times that rate. Its example pins a records-analysis agent to US inference by adding the field to the agent's model configuration.

Anthropic's current data residency documentation treats these as two separate controls. Inference geo decides where model inference runs. Workspace geo, fixed when a workspace is created and currently available only as `us`, decides where data is stored at rest and where processing such as code execution happens. The Managed Agents cookbook says the agent setting covers only where tokens are processed, not what the agent can read or where its data is stored.

The reason to pin an agent is stability. An agent without a pin follows the workspace's default inference geo, and the cookbook notes that a later change to that default reaches sessions that are already running. A pinned value travels with the agent, must be on the workspace's list of allowed geos when the agent is saved, and can still be overridden when a session is created. The documentation also says that attached memory store contents remain stored by Anthropic even when tools run in a self-hosted sandbox, so the pin alone does not settle where every part of a session's data lives.