The biggest obstacle to enterprise AI adoption isn’t model quality. It’s simpler than that. Your legal team has one question before any conversation about automation, efficiency, or use cases can even start: where does the data go?
On November 12, Anthropic opened private beta for self-hosted Claude Managed Agents. Everything runs in your infrastructure now. This deployment option lets organizations run the entire AI agent stack inside their own environment, including the execution sandbox where code runs, the session history agents maintain between tasks, and the credential handling that keeps API keys away from untrusted code. All of it sits behind your firewall, under your policies, in your cloud account or on-premises environment.
This is a meaningful shift. A Claude agent doing something genuinely useful, such as reading internal documents, calling business APIs, writing and running code, or handling customer data, generates a trail of sensitive activity. Under the standard hosted model, that activity happens in Anthropic’s infrastructure. Under the self-hosted model, everything happens in yours. Your security controls apply. Your audit logs capture it. Data you’re not allowed to export stays inside your perimeter.
The architecture is thoughtfully designed for enterprise concerns. Anthropic separates the agent’s execution sandbox from the orchestration loop, which is the process that decides what action the agent takes next based on results from previous tool calls. Both components are also separated from the credential vault where API keys live. Your secrets manager handles credentials rather than exposing them to agent code. Logs flow via OpenTelemetry, an open standard for collecting telemetry data that most existing monitoring stacks can ingest. Network access rules come from your firewall configuration, not from defaults set in someone else’s cloud.
For businesses in healthcare, financial services, defense, or legal services, this changes the conversation with compliance teams. The standard objection to AI agents has been that you can’t prove what data the model saw or where it went. Self-hosted deployment changes that. Everything stays in your environment, so you can prove it.
The same logic applies to smaller professional services firms. Law offices, accounting practices, HR platforms, and healthcare technology companies all handle sensitive information under various confidentiality requirements, and cloud-hosted AI processing has been a nonstarter for many of them. Self-hosted agents change that calculation.
Access is currently by invitation only. You can’t flip a switch on Monday morning. But it is worth putting on your infrastructure roadmap now, and worth bringing to your compliance and security teams before your competitors start building with it.
Want to explore how self-hosted AI agents could benefit your business? Let’s talk.