A process you can argue with
I almost built a publishing pipeline in a workflow builder, then noticed what I'd lose. A node graph gives you nothing to argue with: keep the part you're still reasoning about in the conversation, and push only the settled part down into the loop.
Most automation starts the same way. You open a workflow builder, drag the steps onto a canvas, wire them together, and hit run. n8n, Zapier, the rest. I almost did that tonight for a small publishing pipeline. Then I noticed what I’d lose.
A workflow builder turns your logic into a graph of nodes behind a canvas. Once it runs, the reasoning is sealed inside it. To change how it behaves you stop the run, go back into the builder, and rewire. Building and running are two separate rooms, and the logic lives in the first one.
I built the pipeline another way: a few skills written in plain language and a couple of databases I can see, operated through conversation. One of them was the drafts inbox I’d built the night before. Same mechanics, roughly. The difference showed up the moment I changed my mind. Partway in, I saw that one of the fields I’d added was the wrong kind of thing, a category I didn’t actually want. I said so. It pushed back, we argued it through, the field came out. That happened in the same place the pipeline runs, in the conversation, not in a separate builder I had to reopen.
That’s the distinction, and it’s not about capability. The builder could have done the mechanical work fine. What it can’t do is let me reason with the process while the process is live. A node graph gives you nothing to argue with.
The conversation can carry it because the logic and the state are in the open. The skill is the specification, written as prose I can read. The database is the state, sitting in plain view. Nothing load-bearing is hidden inside a compiled loop. A closed automation keeps its reasoning under the surface; this keeps it on top, which is what makes it something you can think with instead of only configure.
None of this is a case against workflow builders. The moment a process is settled and runs often, unattended, you want the closed deterministic loop, fast and cheap. My pipeline has a piece exactly like that, a nightly job that only executes. The split is the point. Keep the part still being reasoned about in conversation, and push the part that’s settled and repeating down into the loop.
The cost falls the same way. A builder driving a model meters every step out to an API, billed on top of the tool itself. Keeping the reasoning in the conversation keeps it inside a flat subscription. While a process is still moving and low in volume, that’s the saner side to be on, and high volume is exactly when you’d move that part down to the deterministic tier anyway.
People hear “conversational” and picture a chat box bolted onto an app. This is the other thing it can mean. The process itself is the conversation, and you change it by talking, not by opening the builder. Once you’ve run something that way, a canvas full of boxes starts to look like the place your reasoning goes to get locked in.