This talk explores our production-hardened approach to sandboxing agentic workloads using cloud-native container technologies.

We'll walk through our multi-layered isolation strategy:

Container-Based Sandboxing: AI agents operate exclusively within Docker containers where the original codebase is mounted read-only, preventing direct modifications to production code.

Overlay Filesystem Architecture: Overlay mounts to create ephemeral writable layers where agents can safely generate and test code changes. Modified files are extracted over the network, validated before applying to the actual codebase.

Network Segmentation: We implement strict network policies and VPN-based isolation to prevent agents from accessing unauthorized external tools, downloading potentially malicious code, or pivoting to other systems in the infrastructure.

We're currently evaluating Kata Containers for kernel-level isolation and Kata-CC to enable secure agent operations on sensitive confidential workloads.