You want to get some information that might actually change rapidly about the state of the kernel, in the hope that you might be able to get a bit of entropy from it.
The solution is highly operating system-specific. On systems with a /proc filesystem, you can read the contents of all the files in /proc. Otherwise, you can securely invoke commands that have some chance of providing entropy (especially if called infrequently). On Windows, the Performance Data Helper (PDH) API can be used to query much of the same type of information available on Unix systems with a /proc filesystem.
Mix any data you collect, as well as a timestamp taken after the operation completes, into an entropy pool (see Recipe 11.19).
The internal state of an operating system can change quickly, but that does not mean there is necessarily any entropy there to collect. See Recipe 11.19 for a discussion about estimating how much entropy you are getting.
Definitely do not query sources like these very often, because you are unlikely to get additional entropy running in a tight loop, and the overhead involved is extremely high.
On systems with a /proc filesystem, pretty much all of the interesting operating system-specific information you might want to query is available by reading the files in the /proc directory. The contents of the files in that directory are updated as the user reads from those files. Open the files anew every time you want to poll for possible entropy.
On systems without /proc, you can try to get information by running commands that might change frequently and capturing all the data in the command. Be sure to call out to any commands you run in a secure manner, as discussed in Recipe 1.7 and 1.8.
When calling commands, state does not actually change very quickly at all, particularly on systems with few users. It is popular to query the ps and df commands (using the flags that give the most entropy, of course), but there is often almost no entropy in the output they produce.
Other commands that some operating systems may have, where there might be some frequent change (though we would not count on it) include the following:
sysctl: Use the -A flag.
netstat: Use the -s flag if you want to see highly detailed information that may change frequently on machines that see a lot of network traffic.
tcpdump: Ask it to capture a small number of packets.
Often, these commands will need to run with superuser privileges (for example, tcpdump). Depending on your threat model, such commands can possibly be more useful because they're less subject to attacks from local users.
This approach can be a reasonable way of collecting data from the network. However, note that attackers could possibly feed you packets in a predictable manner, designed to reduce the amount of entropy available from this source.
Recipe 1.7, Recipe 1.8, Recipe 11.19