We have one machine that occasionally looses network connection (for whatever reason).
If that happens during a build the server will kill the running build on the server side and as soon as the agent connects again it will start a new scheduled build on that node even though there is already a build active. The agent seams to know about that still running build as it writes to Log when it finally finishes.

I'll attach a log that shows that behavior nicely as there were broken dependency builds in that cycle and so a lot of short builds were triggered while the one build was still running.
As the build was running here at least the wrapper has to know that there are still processes running that it started. In a situation like that the agent must not start new builds.