Persistence and Human-in-the-Loop

You have state. You have a graph. On a large refund you still stop in the same program run.

Real life needs this:

1. Monday — agent reaches “needs approval”
2. Save the state
3. Program exits
4. Tuesday — human says yes
5. Load the state and continue — refund, without looking up the order again as if nothing happened

That is persistence + human-in-the-loop (HITL).

Same refund story

Person’s question: “Please refund order 4412.”
Amount is large. Graph path:

lookup → decide → wait_human

At wait_human, do not pretend the human is inside your while loop.

  run graph → wait_human
       │
       ▼
  SAVE state to a file
       │
       ▼
  (time passes — human decides)
       │
       ▼
  LOAD state + set approved = yes
       │
       ▼
  continue → refund → done

Persistence = remember on disk

Persistence means: write state somewhere that survives when the program ends.

For learning, a small JSON file is enough:

{
  "order_id": "4412",
  "amount_usd": 900,
  "size": "LARGE",
  "approved": "not_yet",
  "status": "needs_approval",
  "next_node": "wait_human"
}

Later you might use a database. The idea stays: state outlives one run.

Also save where you stopped (next_node or status). Otherwise you do not know which node to resume.

Human-in-the-loop = a person is a step

Human-in-the-loop means a person must decide before the workflow continues.

The human is not a model call. Your program:

1. Stops at wait_human
2. Shows the facts (order, amount)
3. Waits for yes/no (button, ticket, or a line in a file for the lab)
4. Writes the decision into state
5. Runs the next node

Worked walk-through

Run A (Monday)

1. lookup → amount 900
2. decide → LARGE
3. wait_human → reply needs approval
4. Save state + next_node: after_approval
5. Exit

Human (Tuesday)
Approves order 4412.

Run B (Tuesday)

1. Load state
2. Set approved: yes
3. Node refund → “Refunded $900…”
4. done

No second random lookup. State already had the amount.

Rules that save pain

What you can build now

Next: Agent Frameworks — tools that package these ideas so you write less runner code yourself.

Words to keep

See it in Code

The Code panel runs the large-refund path, saves state to a file, simulates a human yes, loads state, and finishes the refund. Check the Example console.