One step's answer is the next step's question.
A chain fetches something real, hands it to a model, and delivers the interpretation. Each hop is billed on its own, and a failure late in the chain never makes you pay for the early part twice.
What a chain gives you that a single step does not
A scheduled scrape tells you what a page says today. A scrape plus a model tells you what changed and whether to care — which is usually the thing you actually wanted.
Fetch then interpret
Twenty network-live tools return data. None of them returns a decision. The second step is where the decision comes from.
Diff against yesterday
A prompt can reference the previous run's fetched output, so "what changed" is a question you can just ask.
Up to ten sources
Fan out across a set of pages and get one merged brief, with the sources labelled so nothing is attributed to the wrong one.
One source down is not a failure
With a dozen sources something is always broken. The model is told what was missing rather than quietly implying it read everything.
Priced per hop
The fetch and the model are separate charges, both itemised on the run.
Retries reuse the fetch
A model step that fails is retried against pages already paid for. Durable execution, not a rerun.
Two steps is the whole design
Every researched use case fits in one fetch and one interpretation. A third step multiplies the cost and the failure modes for cases nobody asked for — so when someone does ask, the answer is a new kind of job, not a graph engine nobody can debug.
Start building