An autonomous vehicle must safely navigate a complex set of scenarios under varying driving speeds and in potentially unpredictable environments. This requires the vehicle to adjust its response time deadline in the presence of scenarios that demand faster yet high quality decisions. Choosing such a deadline requires the developers to trade-off between the accuracy of the components being used and their runtime at development time for all possible scenarios that a vehicle may encounter. The choice of a lax deadline may lead to accidents in emergency scenarios, whereas a tight deadline might require the usage of low-accuracy components that could lead to uncomfortable or unsafe maneuvers (e.g. maneuvers arising from detecting obstacles late). In our work, we study the design of software systems that enable AVs to dynamically adapt their end-to-end deadlines to their environment (e.g., driving speed, trajectories of other agents). We are developing ERDOS, an open-source system that exposes explicit time control mechanisms atop a streaming data-flow abstraction. Lastly, to show the benefits of ERDOS, we are developing Pylot, an open-source AV stack that seamlessly runs atop the CARLA simulator and a Lincoln MKZ vehicle. Pylot enables both the research of new AI algorithms and models, and of software systems for AVs.