I think every new kind of "engineering" in AI is just us learning where to put the structure.
First it was prompt engineering: find the right words. Then context engineering: find the right information. Then loops: let the model try, check its work, and try again until it's done. Now people are calling the next step graph engineering.
What is it?
The idea is simple. Instead of one agent running in one big loop, you break the work into nodes and draw edges between them. One node plans, a few nodes do the work in parallel, another node reviews, and the edges decide what happens next. The graph is the program. The model is just what runs inside each box.
Why I like it
Visibility: A loop is a black box. You see the input and the output and hope the middle was fine. A graph forces you to say out loud what the steps actually are. When something breaks, you can point at the node.
Control: Some parts of a task should be creative and some should be boring. A graph lets you keep the model free where that helps, and keep it on rails where it doesn't. The tests have to pass. A human approves the deploy. The reviewer can send it back.
Why I'm unsure
Unfortunately, I think a lot of graphs are just flowcharts with extra steps. It is very fun to draw boxes and arrows. It feels like progress. But every node is another prompt to maintain, another place for context to get lost, and another handoff where one model forgets what the last one was thinking.
The models also keep getting better. A graph you drew to work around a weakness today might just be in the way in six months. Writing the structure down is good. Holding onto it too tightly is not.
So what?
My take for now: start with a loop. Add a node only when you can name the failure it fixes. If you can't name it, you probably don't need the graph, you just like drawing it.
I don't think the real skill is graphs. It's the same one it always was: knowing which parts of a problem need judgement and which parts need rules.