Visual attention is often conceived as a high-speed serial system moving rapidly from one object to another at rates of a few dozen milliseconds per item. We present four experiments demonstrating that this high-speed model is incorrect. Subjects identify two objects, presented at separate times. We measure how long the first object continues to interfere with accuracy on the second, and hence the time-course of the first object's attentional demand. We find interference for a half-second or more-roughly 10 times longer than might be predicted from conventional visual search paradigms. In further experiments, we show that the time-course of interference depends upon the number of attended objects, not the number or complexity of responses. Even objects which require no response, such as nontargets in visual search, can still produce long lasting interference on subsequent identification. We suggest that visual attention is not a high-speed switching mechanism, but instead a sustained state during which representations of relevant objects become available to guide behavior.