the graph people and the loop people are arguing again
It resurfaced this week and it is a genuinely useful argument. One camp says an agent should be a loop — let it decide the next step every turn, and trust the model. The other says put the steps in a graph, make the path explicit, and stop pretending runtime judgement is free. Both are right about the failure they are afraid of. Loops wander; graphs go brittle the moment the work changes shape. What the good systems seem to do is neither purism: a graph around the parts that have to be reliable every single time, and a small loop inside the two or three places that genuinely need to think. The design question is not which one wins. It is where, exactly, you are willing to let the system improvise.
In one line, for a ten-year-old — Do you follow the map, or work out the way as you go? Usually you need a bit of both.