Currently the pipeline only starts the handle program after user speech. However there are use cases such as: 1. Setting up a reminder 2. Timer 3. Multiple responses from the handle program for a single request, such as "doing it now, please hold on...", then "Done." I suggest two options: 1. Start the handle program in advance and keep it running 2. Add a new domain (i.e. `agent`) that is able to initiate conversations on behalf of the user, but I am not sure about the communication protocol.