Hi,
network overhead is dependent on the data you are gathering. Normal metric data does only consume in the order of 100 bytes or less per metric per interval (15s) - if I remember correctly. The more metrics you gather the more data you send. You should keep the number of metrics to 1000-3000 per agent as a best practice.
What consumes more network bandwidth are transaction traces, error and stall events. If you are actively tracing all the time or generating lots of errors that will produce more data. But there are clamps in place (e.g. max 5000 components per trace, max 10 errors per interval) to also step on the breaks if things go south.
We have been running APM in production with 10,000 of agents for years and I have not heard one customer complaining about the network traffic we generate!
Keep your agents healthy and you won't even notice the data they send in the network.
Ciao,
Guenter