Neural circuits encode and transmit information to support a diversity of tasks. Such neural coding is corrupted by noise and uncertainty, which in turn results in different magnitudes of error in the signals that are received by downstream regions. Therefore, even in the same task and given a fixed distribution of input signals, the performance of a neural code is characterized by a distribution of errors. Different task demands might necessitate distributions with different properties; for example, a task that requires avoiding failure at all costs might necessitate coding schemes that minimize the maximum possible error value. Despite the importance of error distributions, current theories of neural coding focus predominantly on optimizing average performance. Here, we generalize this approach to optimize full error distributions. We do so by interpreting error as a random variable whose statistics depend on the input distribution and the system parameters. We show that optimizing for different target error distributions yields different coding schemes and performance tradeoffs. To demonstrate the relevance of this approach, we consider a simple model of decoding escape direction from a neural command signal. In the fly, this computation is performed by two parallel pathways: Descending Neurons, which control a slower but higher-precision escape, and the Giant Fiber, which triggers a rapid but lower-precision escape. We postulate that a rapid escape might require minimizing the probability of exceeding a critical error threshold, and we demonstrate that such a coding scheme differs from one that minimizes average error. We further show that dynamically adapting the code to switch between these schemes would be costly and error prone. Our theory thus provides a candidate explanation as to why escape behavior in the fly is controlled by two parallel pathways that yield different performance characteristics.