Draw the state, not just the nodes
A diagram of prompts and arrows can make an agent workflow look deterministic. The important contract is the state carried across those arrows: input identity, retrieved evidence, current decision, approval status, tool result, and terminal outcome. LangGraph documents checkpointers for thread-level snapshots and stores for information shared beyond one thread. Its interrupt mechanism can persist state while waiting for external input. These are concrete runtime capabilities; they do not decide which business facts are safe to persist or how long they should be retained.
Design for interruption
A worker can stop between proposing a side effect and observing its result. On restart, the graph must know whether the effect happened, whether it is safe to retry, and which checkpoint was authoritative. Model each consequential node with an operation identifier and an explicit status such as proposed, approved, attempted, confirmed, or failed. Do not infer completion merely because a model says it sent a message. The service that owns the side effect should provide a receipt or query path, and the workflow should store the receipt with the state transition.
Replay is a separate feature
Replaying from a checkpoint is useful for debugging, but a second execution can duplicate a payment, ticket, email, or database write if the node is not designed for it. LangGraph's interrupt documentation describes how execution resumes from saved state; it is the application designer's job to protect external effects. For each node, mark whether it is pure computation, a repeatable read, or a write. Where writes are necessary, use idempotency keys or a reconciliation step. A visual graph should display these labels so reviewers can locate the dangerous edges.
Inspect the trace
A useful trace shows the input version, node versions, chosen branch, tool arguments, external receipts, and final state. It should also show where a human decision entered the run. Keep sensitive data out of broad traces and define retention deliberately. When a run fails, compare the last durable checkpoint with the external system before resuming. The editorial test for a graph is simple: can another engineer explain what happened, repeat safe computation, and avoid repeating a side effect? If not, more nodes will make the diagram richer but the system less governable.
What to carry into the work
- Define state fields and retention before drawing edges.
- Label pure, read, and write nodes.
- Give external writes operation IDs and receipts.
- Test interruption and replay with harmless fixtures.
Sources & dates
- Persistence — LangGraph documentation ↗LangChain · Undated source · Checked 16 Sept 2026
- Interrupts — LangGraph documentation ↗LangChain · Undated source · Checked 16 Sept 2026
Unknown source dates stay undated. Preparation is not publication; no historical byline or interview is implied.