Good speech quality with low-delay coding at 8-16 kb/s has been obtained using backward adaptive analysis-by-synthesis codecs, such as Low-Delay CELP (future 16 kb/s standard), Low-Delay Vector Excitation Coding (LD-VXC), and backward adaptive tree/trellis codecs. This paper presents design and performance trade-offs for the low-delay analysis-by-synthesis codecs at rates of 8-16 kb/s. A number of approaches for improving the speech quality at 8 kb/s are discussed. Backward pitch prediction configuration is compared to a closed-loop forward configuration similar to that used in the conventional forward CELP for the adaptive codebook. Finally, the robustness to transmission errors is discussed and a number of trade-offs for reducing the sensitivity to transmission errors are presented.