Estimation for agile processes is a bit of an odd duck. One the one hand it needs to be done before work starts, but on the other hand, the more such stuff is done before work starts, the less agile the process becomes. To overcome this apparent contradiction, several people have tried to devise more “agile” ways of estimating work.
One of these approaches is Planning Poker. See also AGILE IN ACTION: XPDAY2006: Experiments in Agile Estimation