graph-node-and-edge-designlisted

# Graph Engineering: Nodes, Edges, and Shared State **Source**: AI Builder Club, "Graph Engineering Guide (2026)" — describes the pattern as implemented (with different specifics) by LangGraph's `StateGraph`, Microsoft AutoGen's GraphFlow, and Google ADK's graph-based workflow agents, all predating the "graph engineering" terminology itself. ## The three essential components 1. **Nodes** — work units, typically specialized agents or deterministic functions. Real specialization means a distinct prompt, tool set, and clean context per node — not one agent's transcript switching roles mid-conversation (see `loop-vs-graph-decision`'s 4-question test, question 1). 2. **Edges** — routing between nodes: straight (always proceed to the next node), conditional (branch based on a check), fan-out (spawn multiple parallel branches), fan-in (wait for parallel branches and merge). 3. **Shared state** — an object that travels along the edges, carrying task data, notes, drafts, and verdicts, growing as it passes through nodes. ## The canonical starter pattern **Researcher → Writer → Reviewer**, with a conditional edge: pass → Ship; fail → loop back to Writer. The state object grows explicitly at each step: ``` {task, notes} # after Researcher {task, notes, draft} # after Writer {task, notes, draft, verdict} # after Reviewer ``` Note this is small and explicit — every field's origin and purpose is traceable, which is what makes the grap