Host any agent harness | Agent 37
Give every customer an agent in its own isolated sandbox, at up to 95% lower cost than most sandbox providers.
Enterprise AI Infrastructure
Summary
Red Hat demonstrates how Anthropic's self-hosted sandboxes for Claude Managed Agents integrate with OpenShell, an open source sandbox runtime, enabling enterprises to keep code execution on their own infrastructure while outsourcing AI reasoning to Anthropic's cloud.
Highlights
- Anthropic's self-hosted sandboxes split the work: reasoning and orchestration run in Anthropic's cloud while code execution stays on your infrastructure
- OpenShell adds kernel-enforced isolation including Landlock, seccomp filtering, network namespace isolation, and deny-by-default sandboxing to protect the execution layer
- The integration works out of the box on both a developer laptop with Podman and a Red Hat OpenShift AI cluster
- OpenShell is an open source project started by NVIDIA where Red Hat is an active contributor and maintainer
- Self-hosting is ideal when agents need access to data that cannot leave the network, internal services, or must run under organizational audit controls
auto-generated
Context
Audience
Enterprise developers and platform engineers evaluating AI agent deployment strategies, particularly those concerned with data sovereignty and sandbox security
DomainAI Agent Security
Formattechnical blog post
Accessfree online
Source Overview
Anthropic’s self-hosted sandboxes for Claude Managed Agents let enterprises keep code execution and data on their own infrastructure while the model’s reasoning and orchestration run in Anthropic’s cloud.
OpenShell provides the missing security layer for this execution environment with kernel-enforced isolation, granular per-binary network policy, credential isolation, and deny-by-default sandboxing. By integrating Claude’s container-per-session workers with OpenShell sandboxes across Podman and OpenShift drivers, organizations can safely run untrusted agent-generated code on their own systems while returning only sanitized results to Anthropic.
Topics
Related
Claude Managed AgentsOpenShellPodmanOpenShift AILandlockseccomp
