Most workflow engines pass data between steps through one shared, flat state dict. It is convenient at first, but it hides the real data dependencies: any node can read or clobber any key, and a rename in one place silently breaks a consumer somewhere else.

flow takes the opposite stance. Every node declares typed input and output ports, and every edge carries an explicit mapping — nodeA.output.x feeds nodeB.input.a. The wiring is spelled out, not implied by matching key names.

That single decision buys a lot. The graph can be validated before it runs: unknown ports, two producers feeding one input, or an unfed required input all fail fast at load time. Port-local interpolation means a node's prompt template only sees its own inputs, so there is no accidental coupling to unrelated state.

It also makes the workflow legible. A reader (or a generated ASCII diagram) can see exactly which data each edge carries. When you later compile the workflow to Python, the ports become ordinary function parameters — no magic global to thread through.