Day 4: Production Patterns - Persistence and Human-in-the-Loop

Source: Dev.to

Persistence (Memory)

Standard scripts lose all state once they finish. LangGraph provides a built‑in Checkpointer that saves the workflow state to a database (or in‑memory store) after every step, enabling:

• Long‑running conversations – e.g., ChatGPT‑style threads.
• Resume on failure – reload the state after a crash and continue.
• Time travel – jump back to a previous step and explore alternative paths.

import { MemorySaver } from "@langchain/langgraph";

// In production you’d typically use PostgresSaver or RedisSaver
const checkpointer = new MemorySaver();

const app = workflow.compile({ checkpointer });

// Identify the session with a thread_id
const config = { configurable: { thread_id: "session-123" } };

Human‑in‑the‑Loop

Sometimes you need a human to review or edit an agent’s output before it proceeds (e.g., reviewing a drafted tweet).

Interrupting the graph

const app = workflow.compile({
  checkpointer,
  // Stop execution right before the "human_review" node
  interruptBefore: ["human_review"],
});

Execution flow (Mermaid diagram)

graph TD
    Start[__start__] --> Agent
    Agent --> HumanReview[Human Review]
    HumanReview -->|Interrupt| Pause((PAUSE))
    Pause -->|Resume| End[__end__]

Step‑by‑step

Run

The user requests a tweet. The agent drafts it, then the graph pauses at human_review.

Pause

The script exits or yields, leaving the draft in the current state.

Inspect

The UI (or user) fetches the state to display the draft.

const state = await app.getState(config);
console.log("Draft:", state.values.messages.at(-1).content);

Resume

After the user approves or edits the draft, the graph is resumed.

import { Command } from "@langchain/langgraph";

// Resume execution with a command
await app.invoke(
  new Command({ resume: "Approved" }),
  config
);

Full‑stack Agentic Architecture Recap

These patterns form a blueprint for modern AI applications: you’re no longer just prompting models—you’re architecting robust, cognitive systems.