Friendly fire is going to get crazy. Can’t trust an LLM on its own for more than a few iterations..

Don’t worry, it will auto compact its context.