The solution is going to be specific to whichever application framework is being used so a root cause will need investigation.
A common test is to turn off AutoProbe completely in the agent profile - introscope.autoprobe.enable = false
just to check that performance is fine with the agent loaded but no instrumentation happening.
If it is fine with AutoProbe disabled, and you have a lot of time and patience, you can reactivate instrumentation by setting the above property back to true and you can go through toggles-typical.pbd and switch off options one by one to check for any that significantly improve performance.
Indeed if you are already using full instrumentation instead of typical, you can tell this because you will have something saying ....full.pbl instead of typical.pbl, make sure you use the typical pbl.
Also disable any custom instrumentation, that is pbds that you wrote yourself or you added that were not delivered directly with the agent.
After that it is most useful to review transaction traces to see what the application actually does when you 'jump between screens', it gives a picture of the code involved, and then consider disabling instrumentation of metric types that appear there, perhaps Struts or XML processing if that sort of thing is enabled.
As a general rule it would be useful to look for any metrics that have high values of Responses Per Interval or Concurrent Invocations, it could be something like an EJB that is busy processing whatever the screen takes as input.
If there are calls to external resources through Sockets, there can be a slowdown at this point so you can try disabling Socket Tracing.
We can go on and on suggesting what it could be..
But you should start by reviewing the transaction traces to get a picture of what a jump between screens actually means in code terms.