State Machines & Graph Agents — The Agentic Control Loop

Modeling Agents as Graphs

The most reliable production agents are not free-form loops - they are state machines modeled as directed graphs. This is the core insight behind LangGraph and similar frameworks.

Why Graphs Beat While-Loops

Anatomy of a Graph Agent

// LangGraph v1.x Structure (Annotation.Root + START/END + interrupt):
const AgentState = Annotation.Root({
  messages: Annotation<string[]>(), plan: Annotation<string>(), quality: Annotation<number>()
});

const graph = new StateGraph(AgentState);
graph.addNode("planner",    plannerAgent);  // Creates a plan
graph.addNode("executor",   executorAgent); // Executes plan steps  
graph.addNode("reviewer",   reviewerAgent); // Reviews output quality
graph.addNode("human_gate", () => ({ approved: interrupt("Approve plan?") })); // Dynamic interrupt()

// Edges define the flow:
graph.addEdge(START,      "planner");
graph.addEdge("planner",  "executor");
graph.addEdge("executor", "reviewer");
graph.addConditionalEdges("reviewer", (state) => {
  if (state.quality >= 0.8) return "human_gate";
  return "executor"; // Loop back for another attempt
});
graph.addEdge("human_gate", END);
const app = graph.compile({ checkpointer: new InMemorySaver() }); // PostgresSaver in production

LangGraph v1.x & LangChain v1 Standards: For standard tool-calling loops without custom topologies, import create_agent from langchain.agents (which replaces the deprecated langgraph.prebuilt.create_react_agent). For custom state machines requiring Human-in-the-Loop pauses, call interrupt(payload) dynamically inside any node (replacing legacy NodeInterrupt and static compiler breakpoints) and resume execution via Command(resume=value) backed by InMemorySaver (local dev) or PostgresSaver (production).

Key Concepts

• Nodes: Individual processing units - can be LLM calls, tool executions, or pure functions.
• Edges: Connections between nodes. Can be unconditional (always follow) or conditional (branch based on state).
• State: A shared data structure (often a TypedDict or Pydantic model) that flows through the graph.
• Checkpoints: Snapshots of state at each node - enables time-travel debugging, persistence, and resumption.

Common Graph Patterns