A bunch of folks are hyping loop and graph as the new trending AI keywords. I tried implementing some of them myself to unpack these vague keywords. I think I can give you a basic understanding of what they mean, starting with what you're all familiar with, a prompt, then adding one idea at a time until we get to a graph.
My original sketch. Top to bottom: prompt, harness, loop, then graph. Each step is the previous one plus one idea.
Start from a prompt
Prompt. You write the question, attach context like RAG, files, images etc and then model answers. You do everything in the question box e.g., gather context, decide what to ask, judge the result, decide what to do next
Harness. Tools like Claude Code or Cursor wrap the same model call and add one important thing: tool calls. The model can now "do" things like edit a file, run a command, and read the output, instead of telling you what to do. You still prompt, but you stop hand-feeding them like before.
Loop: we let model handle the prompt itself
To me, this is what people really mean by "loop". The harness is the same. The only new thing is who writes the next prompt. It is generated from the result of the last iteration, instead of you typing it.
Sounds like a small tweak but it is not. Writing the next prompt is deciding what matters about what just happened. Once the system does that, you leave the inner cycle and focus on setting the goal.
Graph: loops with a manager/orchestrator
"Graph" is the most hand-wavy one, so it takes the whole bottom half of my sketch. A planner splits the goal into sub-goals. Each sub-goal runs in its own harness with a narrow job. Results consolidate. Tests run, also as harness runs, and a failed test sends work back to its goal. Then final check: is the overall goal met? Yes ends the run. No sends the plan back to the planner with what it learned so far.
What the words mean to me now
Each step from prompt to graph hands the system one more task that used to be ours. Harness takes the doing. Loop takes writing the next prompt. Graph takes deciding what happens next. Still, whether you become a proxy who just passes prompts along or a driver who steers depends on your intent.