I tweaked it slightly yesterday so it is now (# cores + 2 tasks) allowed 'in-progress' for the first 3 hours. This will allow the client to be able to slightly anticipate finishing a job and request more work so that cores should not go idle.

And to answer your question. The scheduler first handles results that are reported complete and then assigns new work for the device. As an example, assume a computer with 8 cores and currently has 10 jobs locally, 2 of which are complete. Then A) the computer will be running with all 8 cores able to be working and B) when it goes to request more work it will report the 2 complete tasks and then be able to download up to 2 more tasks during the same request.