Durable HITL with a Postgres Checkpointer

Your human-in-the-loop agent silently loses its state when the worker restarts. Here's the 10-line fix, a Postgres checkpointer that survives the crash.

Ships ~June 10, 2026

Lessons

The silent failure
Watch a human-in-the-loop agent lose its interrupt state the moment the worker restarts, and why in-memory state is the trap.
What a checkpointer persists
The LangGraph checkpointer interface: what gets saved, when, and how a thread resumes exactly where it paused.
The 10-line fix
Swap the in-memory saver for a Postgres checkpointer, the whole change, wired and explained.
Proving durability
Kill the worker mid-interrupt, restart it, and watch the agent resume the human approval it was waiting on.