The Agent resource consumption depends on a lot of variables. Things such as the amount of specific code executions (custom sensors) and how sensors are configured such as capturing exception stack traces and how much db query length is captured. In production environments the agent is configured to consume no more than 5% of the CPU of the system under diagnosis. It becomes difficult to get a direct comparison between instrumented vs non instrumented, since without the agent we cannot really get the information we need.
Either way there are steps that can be taken (especially in PROD environments) to reduce overhead even further. I recommend you check out this article it outlines a bunch of good best practices to further reduce overhead. Hope this helps.