When an agent reads a file, its execution environment, session history, and model request can each sit in a different place.
Anthropic opened access requests for self-hosted Claude Managed Agents on November 12. Customers can run the API service and agent loop, keep session history in their own database, and host the sandboxes where tools execute. Anthropic directs interested customers to their account team. Its documentation lists PostgreSQL 16 or later for persistent storage.
The earlier self-hosted sandbox option moves tool execution into the customer’s environment while Anthropic still operates the service. With the new option, the customer also runs the service and installs upgrades. That changes who operates the infrastructure and stores the session history.
Trace one tool result through the deployment
Use a synthetic invoice in a test workspace. Mark it with a string unique to that run and ask the agent to read it. Follow the result from the sandbox into the session database, then check the outgoing request to the inference provider. Anthropic’s documentation says those requests carry the conversation so far, including tool inputs and results. The default is Anthropic’s API. Customers can also configure a supported cloud provider or their own gateway.
That distinction matters when a policy permits a company to retain session history but also restricts which services can receive document contents. Review both controls before sending real files. The test above is a proposed deployment check based on Anthropic’s documentation.
Abdessalam Alaoui
I'm Abdessalam Alaoui, founder of Musthave.AI. I write practical guides to AI models, tools, generative media.
