feat: rl tuning for idealized react agent #405
Description
The principled way to do react is to force the model to take each individual step in order: Observe -> Think -> Act -> Continue until Done. Our current react problem solving framework isn't able to do that because the model tends to call tools too early and skip thinking, etc...
To test our rl tuning on mellea programs, we could try to train a model to act through this idealized react loop; and see if the model is more performant.
As a part of this, we would have to write a Mellea program that actually looks like the react loop.