There‘s many approaches being discussed and it will depend on the size of the task. You could just review a plan and assume the output is correct but you need at least behavioural tests to understand what was built fulfilled the requirements. You can split the plan further and further until the changes are small enough to be reviewable. Where I don’t see the benefit is in asking an agent to generate test as it tends to generate many useless unit tests that make reviewing more cumbersome. Writing the tests yourself (or defining them and letting an agent write the code) and not letting implementation agents change the tests is also something worth trying.
The truth is we’re all still experimenting and shovels of all sizes and forms are being built.
Would you say you’re able to draw a diagram of the application architecture out of your head or do you treat it as a black box? Do you need an AI to debug issues or not? In my experience with spec driven development, even if reviewing every single PR, it is hard to develop a mental model of the codebase structure unless you invest on it. It might be fine to treat it as a black box, not arguing the opposite but will all software be a black box in the future?
ElevenLabs and Gemelo.AI are services that both support text input streaming for exactly this use-case. I am not aware of any open-source Incremental TTS (this is the term used in research afaik) model but you can already achieve somthing similar by buffering tokens and sending them to the TTS model on punctuation characters.
The truth is we’re all still experimenting and shovels of all sizes and forms are being built.