my agent got way better when I made it explain its plan before acting

one line asking it to state the plan first and the quality of the actions jumped. it is like the model needs to commit to an approach before it stops flailing.