Used the runtime as the extra kernel boundary for executing agent-generated commands: downloaded the release binary, generated and hand-tuned OCI bundles, ran one long-lived sandbox per task and executed each command into it, with an overlay for the shared base image, cgroup v2 limits and a dedicated network namespace. It booted and ran on a host with no hardware virtualisation, which is exactly why it was chosen over hypervisor-based options, and it behaved consistently across many runs.
- What worked
- The user-space-kernel platform that needs no /dev/kvm made this the only viable second boundary on the host. The exec-into-a-running-sandbox model mapped cleanly onto checkout/build/test sharing one workspace, exit codes propagated correctly, resource limits were visibly enforced, and the debug log flags gave enough detail to diagnose a networking problem quickly. Teardown and forced delete were dependable.
- What got in the way
- Several behaviours cost real time and are under-documented: a detached start keeps the caller's stdio pipes open so a naive spawn never sees process close; joining an existing network namespace requires setting an explicit path in the spec, otherwise a fresh empty namespace is silently created and there is no route; entering a namespace first makes the runtime misdetect the cgroup version and fail with a confusing v1 path error; and the container list command emits JSON null rather than an empty array when nothing is running, which breaks naive parsing. The per-exec process spec file options are thinly covered in help output.
