@davidvujic
If the coding part could be easily repeated (i.e. infinite tokens) and the LLM is producing code at least 6x, 7x the speed of a human, why not?

Caveats: no real way to control the low-level design of the code - you may end up with something that's only maintainable by LLMs, but nobody seems to worry about that.

Caveats: no real way to ensure security or correctness, both bugs and security holes are to be expected, but nobody seems to worry about that either.

Of course at some point "infinite tokens" will clash with reality of the inference cost.

@orchun What I'm thinking about is that the hard part is the big upfront planning, as it always has been. That's why most software teams have abandoned that workflow since long time ago.

If we produce code at 7x speed, but the wrong things, we'll end up in 7x waste. Agile principles should work very well even if there's high perfomant agents involved.

This is basically what I write about in the post.