AlphaEvolve also co-evolves the task prompt.
We have implemented this, but there are many ways to implement this and there might be a drift of context if left completely to the LLM.
Design different ways of adapting the task prompt, analyse (deeply) what is more effective for several scenarios.