More executors than CPUs run the same work slower, not faster. These docker stats peaks come from dk-inside and k8s-pod on this shared 4-CPU machine, so compare ratios rather than absolutes:
| Container | Peak CPU | Peak memory | Limit set |
|---|---|---|---|
| l2-jenkins (controller) | 44% | 987 MiB | none |
| l2-agent-ssh (running inside) | 32% | 143 MiB | 2 CPUs, 2 GiB |
| node:24-slim build container | 98% | 59 MiB | 1 CPU, 1 GiB |
| l2-agents-control-plane (kind node and pod) | 196% | 1,220 MiB | none |
The kind node is the expensive part. This controller keeps zero executors on the built-in node and one per CPU an agent may use (agent-ssh has two and --cpus 2), limits every build container and pod so a runaway test is killed instead of swapping the host, caps concurrent pods, and stops what is idle: the kind cluster, agent-ws and the dockerfile image were removed once this section's builds were done.