Define each number
CPU use needs an interval and a denominator. Memory should distinguish the process, its service group and the whole host. Network byte counters need a direction, interface and time delta. A dashboard that mixes these scopes can make normal cache growth look like a leak or make the CPU use of a multi-process service appear smaller than it is.
Give each agent its own identity
Two machines should not submit data with the same node identity. Reusing a token can overwrite or interleave observations and make an unhealthy server appear healthy. Store tokens in a restricted environment file read by the service manager. Check that the agent reconnects after a restart without opening an unnecessary public management listener.
Keep an evidence trail
A green online badge proves recent contact, not every resource calculation. Compare a few dashboard observations with local counters while the machine is idle and while it is doing useful work. Check service restart counts and logs for reconnect loops. Limit log retention and avoid collecting complete application URLs when only hostnames and byte counters are needed to answer the performance question.
At the terminal
systemctl show example.service -p ActiveState -p MainPID -p NRestarts -p MemoryCurrent
cat /proc/loadavg
ip -s link show