Insects routinely encounter more sensory information than can be translated into simultaneous action. Yet studies of selective behavior are often partitioned into separate literatures on multisensory integration, stimulus salience, physiological state, learning, neuromodulation, and motor control. This fragmentation makes “attention” difficult to define mechanistically and encourages either overly narrow circuit interpretations or broad cognitive labels that exceed the evidence. This article develops a proposed behavioral-neuroscience architecture in which attention is treated as dynamic allocation of processing priority rather than as a single conserved center or unitary faculty. The framework distinguishes competition among external cues from modulation by physiological state and experience, separates sensory filtering from learned valuation, and places action selection within the architecture rather than treating movement as a passive endpoint. Evidence from mosquitoes, flies, honey bees, and bumblebees shows that cue reliability, hunger, reproductive state, threat, sleep, associative history, memory comparison, and neuromodulatory signals can each alter which information influences behavior. These effects occur at different neural and behavioral levels and cannot be collapsed into one mechanism. The proposed architecture therefore emphasizes conditional weighting, recurrent sensorimotor feedback, and explicit taxonomic and methodological boundaries. It generates testable predictions for experiments that manipulate cue reliability, internal state, experience, and action opportunity factorially while measuring both sensory processing and behavior.
