A few weeks ago, the AI world declared that “loops” were out and “graphs” were the future. Forty-one days earlier, it had declared loops the future. For the exact same reason.
I noticed this because of a Substack post that laid the timeline out plainly. One of the big names in AI, the guy who created OpenClaw, spent June popularizing the idea that everyone was just doing “loops” now. Forty-one days later, in July, a different well-known voice was asking whether the industry had already moved on to “graphs.” Two terms, six weeks apart, both describing more or less the same thing.
I don’t think that’s a coincidence, and I don’t think it’s really about AI. My hypothesis: a lot of people are trying to coin the term that goes viral. We’ve seen this before. The “situational awareness” essay that launched a fund of the same name that has since imploded ended up launching a multi billion $ fund. Once one move like that works, everyone wants to be the next person who names the thing. So the pace at which “the next big AI term” shows up isn’t really a measure of how fast the underlying technology is changing. It’s closer to a measure of how many people are trying to get noticed.
Workflow, loop, graph — in plain English
These three words describe the same idea at increasing levels of complexity, and none of them are new or AI-specific.
A workflow is just steps in order. Do A, then B, then C. You’ve built a hundred of these — a checklist, a process flow, an approval chain. Nothing repeats, nothing branches, you just move forward.
A loop is a workflow that repeats until a condition is met. Picture a month-end reconciliation: pull a number from one source, pull a number from another, compare them at step C. If they don’t tie out, go back and get more detail, and keep going until they do. That’s a loop — a workflow with a condition built in.
A graph is a workflow, or a loop, with branches drawn in. Take a procurement example: a purchase order goes to someone for approval. Above a certain dollar amount, it routes to the CFO. Below that amount, it just goes to the requester’s manager. Then it gets paid. That’s still fundamentally an approval chain — it just has decision points added to it.
Here’s the invoice version of the same idea, since it’s the cleanest way to see it side by side. Workflow: one signer, then the next, then the next. Loop: kicked back to the first approver until the dollar amounts match, the way a reconciliation cycles until it ties out. Graph: under $500 skips the second signature, over $50,000 needs two signatures in parallel — same three approvers, just with branch points added.
And here’s the detail that actually demystifies the whole cycle: a loop is already a graph. It’s just a graph whose path happens to circle back on itself instead of forking forward. Graph engineering isn’t a new discipline that replaced loop engineering. It’s a more general word for the same shape.
The one real thing that did change
I don’t want to pretend nothing happened here, because something small and real did. To support more of this graph-style thinking, Claude Code shipped “dynamic workflows” — a feature where the AI writes its own step-by-step plan, instead of a person writing that plan in advance. With loop engineering, you’d hand the AI its inputs and the condition to hit, and it would just keep running in order. With graph engineering, the idea is the AI designs its own steps, parameters, and branches, and runs the whole thing until a condition is met.
That’s worth knowing if you’re building AI tools yourself. It is not evidence that graphs are more reliable than loops, and it’s not a reason to feel behind if you hadn’t heard the term yet.
Key takeaways
- A workflow is steps in sequence. A loop is a workflow that repeats until a condition’s met. A graph is a workflow, or a loop, with branches. Same idea, increasing complexity.
- A loop is technically already a graph — one that circles back on itself. Graph didn’t replace loop.
- The two terms landed six weeks apart, from two different commentators, following the same pattern as prompt engineering, context engineering, and harness engineering before them. That pace is the story, not a research breakthrough.
- The one real product change — an AI that writes its own plan — is a detail for people building tools, not a decision point for the rest of us.
- When the next term shows up on your feed, you don’t need to chase it. Translate it back to steps, repeats, or branches, and move on with your day.
Want the CPE credit? Take the full lesson on EverydayCPE and earn 0.2 CPE credits: workflow-loop-or-graph
Leave a Reply