Currently agent sends the log line per line if the next log message does not arrive to the input channel before a append operation is done to a list which is very unlikely. This causes enormous amount of calls to the scheduler's appendtoagentlog RPC API. Possibly one reason why the scheduler sits at 100% cpu even if only 10 agent are running. Instead optional buffering it could be forced for time/lines to avoid sending 1000 calls per second if the log is proceeding fast.
Easy reproduce is running only the licheck and seeing what happens during it's run.