When eager execution is enabled, you no longer need to worry about graphs: operations are executed immediately. The upshot is that eager execution lets you implement dynamic models, like recursive NNs, using Python control flow. We've published some example implementations of such models on Github:
https://github.com/tensorflow/tensorflow/tree/master/tensorf...
I'd be happy to answer other questions about eager execution, and feedback is welcome.
EDIT: Just because you don't have to worry about graphs doesn't mean that graph construction and eager execution aren't related; take a look at our research blog post for more information if you're curious about the ways in which they relate to each other (https://research.googleblog.com/2017/10/eager-execution-impe...).
