* The automagical guessing of missing estimates. Propagating the estimate for a subtask to other subtasks at the same level makes a lot of sense for a first guess, if I'm any good at task breakdown. The propagating up to the "aunt/uncle" tasks (i.e. siblings of the parent) is a bit stranger - seems like it should dramatically underestimate most top-level tasks, and I'd worry about the anchoring effect. Still, it makes it very convenient and quick.
Have you thought about visually distinguishing autoguessed estimates from manually entered ones? At least for leaf tasks, I'd like a reminder of what I have left to estimate. (You could propagate up the "degree of guessedness" to parent nodes, but you probably wouldn't want to, or it would undermine the usefulness of guessing in the first place.)
* The visual tree structure is nice. I like how the horizontal arrangement puts a natural limit on how deeply you can nest, and therefore discourages overplanning :)