The same approach worked for us with agent runs. We gave them fewer tools up front and let the models just pull what they need