# Kata Containers (/docs/foundation/components/kata-containers)



## Component Category [#component-category]

Cluster operations / sandboxed container runtime

## Component Description [#component-description]

Kata Containers runs each pod inside its own lightweight virtual machine, with its own guest kernel, while Kubernetes manages it like any other pod. BullSequana AI installs it with the `kata-deploy` chart and uses it for agent sandbox pods only.

## Why It Is Used [#why-it-is-used]

In BullSequana AI Foundation, Kata Containers keeps agent-generated code off the host kernel. A sandbox pod that the agent runs commands in cannot reach the node or other pods through a shared kernel, which a standard container runtime does not prevent.

## Learn More [#learn-more]

* [Kata Containers documentation](https://katacontainers.io/docs/)
* [kata-containers/kata-containers on GitHub](https://github.com/kata-containers/kata-containers)

## Deployment notes [#deployment-notes]

Kata Containers is off by default. `KATA_DEPLOY_COMPONENT_ENABLED=true` turns it on, together with the agent sandbox components; see [Agent sandbox](/docs/deployment/configuration-model#agent-sandbox).

When enabled, `kata-deploy` 4.2 runs as a DaemonSet in `kube-system` at sync wave 5 in the common tier. It installs the QEMU shim only on nodes labelled `workload=sandbox` and tolerates the matching `NoSchedule` taint. A DaemonSet, rather than a one-time job, also covers nodes that join the sandbox pool later. The installer is a privileged node workload, so only sites with a sandbox node pool enable it.

The platform owns the `kata-qemu` RuntimeClass. It pins Kata pods to the sandbox nodes and charges a fixed overhead of `250m` CPU and `160Mi` memory for each pod's virtual machine, which the scheduler and the sandbox namespace quota count on top of the container requests.

Sandbox nodes need hardware virtualization: `/dev/kvm` on bare metal, or nested virtualization on cloud VMs. See [Enable and operate agent code sandboxes](/docs/deployment/playbooks/agent-sandbox) for the node pool, image mirroring, network, and verification steps.

## Interacts With [#interacts-with]

* `Agent sandbox`, whose workspace pods run in the `kata-qemu` RuntimeClass in the `coreai-sandboxes` namespace.
* `BSQAI API`, which creates sandbox sessions for agents on top of those pods.
* `Node taints and affinities`, through the `workload=sandbox` label and taint that reserve the sandbox node pool.
