To fix this, we're separating the planning and execution into two steps: – plan what to do
– execute on each step This usually leads to more calls to the language model… but helps it stay on track for more complex tasks

→ View original post on X — @hwchase17