In this paper, we consider a cognitive radio (CR) system consisting of a primary user (PU) and a pair of secondary user transmitter (SUtx) and secondary user receiver (SUrx). The SUtx is equipped with a reconfigurable antenna (RA) which divides the angular space into M sectors. The RA chooses one sector among M sectors for its data transmission to SUrx. The SUtx first senses the channel and monitors the activity of PU for a duration of T_{sen} seconds. We refer to this period as channel sensing phase. Depending on the outcome of this phase, SUtx stays in this phase or enters the next phase, which we refer to as transmission phase. The transmission phase itself consists of two phases: channel training phase followed by data transmission phase. During the former phase, SUtx sends pilot symbols to enable channel training and estimation at SUrx. The SUrx selects the best beam (sector) for data transmission and feeds back the index of the selected beam as well as its corresponding channel gain. We also derive the probability of determining the true beam and take into account this probability in our system design. During the latter phase, SUtx sends data symbols to SUrx over the selected beam with constant power Φ if the gain corresponding to the selected beam is bigger than the threshold ζ. We find the optimal channel sensing duration T_{sen}, the optimal power level Φ and a optimal threshold ζ, such that the ergodic capacity of CR system is maximized, subject to average interference and power constraints. In addition, we derive closed form expressions for outage and symbol error probabilities of our CR system.