Data at work
An AI agent can run on your server and still send customer data out
Anthropic’s new private beta keeps the agent’s work history on company servers. A delivery test can show what still reaches the model provider.

Anthropic opened a private beta on 12 November that lets companies run Claude Managed Agents on their own servers. An AI agent is software that can take several steps towards a goal, such as reading a file, checking a record and drafting a reply. The beta lets a company keep the agent’s work history in a database it controls. The service that runs the model can still receive the conversation, instructions and information the agent’s tools return.
A distributor sorting late deliveries needs to check that next request. The agent might use a tool to read an order on a company server. The next request to the model could include the customer’s address, even though the original order stayed on the company server. Anthropic’s data and access documentation says those model calls follow the organisation’s agreement with the provider that receives them. Access to the beta is by request through an Anthropic account team. The announcement does not give a separate rollout date for Jamaica.
Check what the model receives
Run the first test with your technical lead. Give them two sample deliveries: A104 is late and needs a call, while A105 arrived on time. Put the marker TEST-ADDRESS-47 in the address field of both records. Ask the agent to return just the order IDs that need a call and the reason. The expected answer is A104, late. Ask the lead to check the request sent to the model as well as the reply. If either contains the address marker, keep real addresses out of the pilot and remove that field before the agent can read it. This check covers one field in two sample records. Repeat the check for other fields and exceptions.
Find out who can pause the trial
The AI Agent Readiness Score gives the person planning the pilot a way to discuss the work with operations and IT. Select Score My Readiness and answer the twelve questions for one job, including how people review actions and how they stop the system. Select View Results Without Sharing Contact Details to see the score and the results by area. A practice run with key workflows documented and the lowest option on all other questions returned 6/100 and 0/9 for oversight and guardrails. Those numbers describe the answers entered. They cannot verify what a provider receives. Start with the low oversight score: name the person who can pause the trial. Have them watch the sample test before anyone connects real orders.
Sources and further reading
- Anthropic announces self-hosted Claude Managed Agents
- Anthropic explains data and access in self-hosted Managed Agents
- Anthropic compares hosted and self-hosted Managed Agents
- AI Agent Readiness Score
Sources checked 13 November 2026.