What LangGraph feature allows the agent loop to cycle back to an earlier node rather than terminating?

LangGraph models an agent as a **graph** of nodes connected by edges. A node is a unit of work, usually the model reasoning step or a tool call. An edge says which node runs after the current one finishes. There are two kinds of edge. A normal edge is fixed: node A always goes to node B. A conditional edge is a small function that reads the shared state and returns the name of the next node. Because that function can return the name of any node, it can return a node that already ran. That is the whole trick behind the loop. The reasoning node runs, then a conditional edge checks the state. If the model asked for a tool, the edge routes to the tool node, which then routes back to the reasoning node. If the model gave a final answer, the edge routes to the end instead. The wrong answers describe real features that do something else. Parallel edges fan out to run nodes at once. Memory and checkpoints store state. None of those decide which node runs next, so none of them create the cycle. The loop lives in the conditional edge, because that is the only part of the graph that looks at state and points backward.

The answer is **conditional edges**. In LangGraph you build a StateGraph: nodes are functions that read and update a shared typed state, and edges decide which node runs next. A static edge is unconditional. It always sends control from one fixed node to another. A graph built only of static edges is a straight line that runs once and stops. Conditional edges are different. You attach a router function that receives the current state and returns the name of the next node. That router can return any node name, including a node that already executed. The classic agent loop wires the reasoner to a conditional edge. The router checks whether the last model message contains a tool call. If yes, it routes to the tool node, which has a static edge back to the reasoner. If no, it routes to the end. That back-routing is the cycle. It turns a directed acyclic graph into a directed graph with a loop. The distractors all name real LangGraph features, but none of them choose the next node. Checkpointing persists state so a run can pause and resume. Parallel edges fan out. Neither one routes control flow backward, so neither one creates the loop.