What is new about self-hosted Managed Agents?
Managed Agents is a suite of composable APIs for autonomous agents. The service provides the agent harness, sandboxes, credential handling and session history so long-running agents can go into production. Anthropic launched it as a hosted service and later added self-hosted sandboxes for teams whose code must stay in their own environment.
With the new variant, every part of the service runs in your own cloud or on-premises. Companies operate the Managed Agents server, its database and the workers where agents run their shell and file tools. Developers still use the same core APIs as with the hosted service.
Where do data and model requests go?
Model requests can go directly to Anthropic, to a cloud provider (Amazon Bedrock, Google's Agent Platform or Microsoft Foundry) or to your own gateway. Session history is stored in your own database. If an organization has zero data retention in place with Anthropic, it applies to the requests sent to Claude like any other Messages API call.
For companies in the EU, this matters because they decide where session data and logs are stored. The model requests themselves, however, still leave your environment toward the chosen provider. Check which route and region processes the requests and document this in your data protection and security documentation.
Which security principles apply?
Anthropic names four principles that apply regardless of the operating model. Because unattended agents are not checked by a person at every step, the limits are meant to be built into the architecture.
- Brain and hands are split: the harness calls Claude, code runs in the sandbox; agent loop, history and secrets stay outside.
- Credentials stay out of the sandbox: a proxy outside adds them for MCP tools and supported API calls; when self-hosted, also from your own secrets manager.
- Network access can be limited: an allowlist of hosts per environment; all other requests are blocked.
- Full records: every step including tool calls is recorded and exported over OpenTelemetry.
How do the Managed Agents variants differ?
There are now three operating models: Anthropic hosts everything, Anthropic hosts the service and you host the sandbox, or you host the entire service. The more you operate yourself, the more control you have over data and network, but also the more operational effort for updates, database, scaling and monitoring.
Anthropic itself names typical uses: fixing broken builds, testing releases, triaging alerts and working through bug backlogs. These agents work unattended on your own code and need access to internal systems. That is where the self-hosted variant pays off if internal policies require credentials and logs to stay in-house.
How should companies proceed now?
The feature is a private beta; the announcement does not state prices, regional availability or a timeline for general availability. If you are interested, clarify your requirements before contacting the account team.
- Justify why the entire service, not just the sandbox, must be self-operated.
- Assign operational responsibility for server, database and workers.
- Define allowed hosts per agent environment, for example only the git host and package registry.
- Connect OpenTelemetry records to existing monitoring and set retention periods.
- Record route and region of model requests in the data protection documentation.
Keep it verifiable




