Maybe I didn't convey the idea completely. A system is incomplete until it can test itself.
For example:
There was once AlphaGo, which dominated its opponents, but learned on human data.
Then, over time, there was AlphaZero, which learned exclusively by playing against itself.
The main goal is to allow the agent to learn by playing with itself, solving problems using only the tools at its disposal.
It must be able to write tests that will allow it to self-optimize.