The advanced state of agent software and computing hardware makes it possible to construct complex agents and robots with multiple streams of input such as vision, speech, gestures and data. Such agents, like people (who also have access to multiple input streams), need to effectively manage the input in order to process important information within useful time bounds. This paper discusses processes and architectural components that are used to manage input data. In addition to reduced processing load, input management may also enable symbol grounding. However, some effects are not beneficial. For example, the agent will lack a full accounting of all input data, which means that standard explanation techniques will not function correctly. We propose several techniques for overcoming the disadvantages of input management.