This is inconsistent with our cpulimit policy, for which we are stipulating both the agent and the user process are bound to the same CPU limit and must share the resources.

In this way, it would be best if we'd also account the CPU usage from the agent + the user process on the final statistics output.

One must be careful as the agent is called in two different contexts (beat.scheduler iodaemon or beat.cmdline exp run), but the accounting must always come out right.