To what extent are shifts of attention driven by encoding of visual-spatial landmarks, associated with useful locations, or by encoding of environmental cues that act as symbolic representations, providing information about where to look next? In Experiment 1 we found that when cues were presented with a long exposure time (300 ms) attention shifts were driven by the symbolic identity of cue stimuli, independently of their visual-spatial (landmark) features; but when cues were exposed very briefly, (66 ms), attention shifts were independent of symbolic information, and were driven instead by visual landmark features. This unexpected finding was interpreted in terms of the transient and sustained response characteristics of the M-cell and P-cell inputs to the dorsal and ventral visual streams, respectively, and informed our theoretical proposal that attentional effects elicited by visual-spatial landmarks may be driven by dorsal stream (“where pathway”) encoding; while attentional effects driven by the symbolic identity of cues may be driven by ventral stream (“what pathway”) encoding. Detailed predictions derived from this proposal, and based on distinct physiological properties of the 2 visual streams were tested and confirmed in Experiments 2–6. Our results suggest that a 2-process view of attention shifting can be integrated with dual-stream models of vision. According to this unified theory: (a) Landmarks associated with visually useful locations elicit rapid, nonconscious shifts of attention, via nonsemantic, dorsal visual stream encoding of their features and spatial relationships; (b) Slower, endogenous shifts of attention are elicited by ventral visual stream encoding of symbolic-semantic information.