What does a workflow graph add?
A graph lays out the steps and what can happen next. LangGraph can save workflow state at checkpoints, so the app can return to the same run later when persistence is configured.
Draw “read draft,” “propose a title,” and “wait for editor” on paper. Put a separate path for “editor rejected” and “editor approved.” The graph should make it easy to see when a change can be written.
What runs again when a paused workflow resumes?
An interrupt can pause for outside input. When the workflow resumes, the interrupted step starts again, so work before the pause in that step can run again.
Keep the step that waits for approval free of writes. Put the write after the decision and protect it against a repeat. On paper, send the same fictional decision twice and verify that the final title changes once.
What should you keep before moving or restarting the app?
Choose persistent storage for progress that must survive a restart. The app needs to associate the saved run with the person who started it and the reviewer and recheck permissions when it resumes.
Keep the source draft, approval record and final saved result separately from a checkpoint. If you move hosts or change how state is stored, test a disposable saved run before assuming that pending decisions still work.
What is the difference between the library and hosted services?
LangGraph’s open-source library is MIT-licensed. You can run the library in your own application. Hosting, storage, model requests and optional LangSmith services are separate choices.
Estimate the model calls, saved progress and monitoring you need. A workflow that follows simple rules can run those steps without asking a model each time.