Do meta prompting more. Less complex setup and orchestration. Rely more on self reflect capacity of the models.
That’s it.