We tested LangGraph checkpointing—and kept our current system
A framework can solve a real technical problem and still be the wrong dependency today. We compared one interrupted workflow before deciding whether to adopt it.
The question behind the work
The problem
Interrupted AI workflows need durable state. Our current approach writes explicit JSON checkpoints, while LangGraph offers graph execution, interrupts and checkpointing as a component. The question was whether that extra machinery produced enough practical value at our current scale.
01
What we tried
- Built the same small synthetic workflow twice: once with explicit JSON state and once with LangGraph plus a local SQLite checkpointer.
- Started, interrupted and resumed each implementation in separate processes.
- Repeated the resume after completion to verify that the single prepared effect was not duplicated.
- Kept authority and the simulated human gate outside both implementations.
02
What happened
- Both approaches resumed and completed with exactly one effect.
- LangGraph created finer automatic checkpoints and a useful interrupt cursor.
- The LangGraph version required more source code and an isolated environment containing 38 packages; the native version used the Python standard library.
- One successful synthetic comparison did not demonstrate a material reliability advantage.
03
What we learned
Checkpointing mechanics and operating authority are different concerns. A framework can manage execution state, but it does not supply our permission model.
Automatic checkpoints are attractive, yet dependency and maintenance costs must be justified by repeated real failures—not by architectural elegance.
A good Build-vs-Adopt decision can end with learning a pattern instead of installing a component.
04 · Outcome
The decision
KEEP CURRENT. Retain the file-native approach and the useful patterns: checkpoint material transitions, preserve a stable interrupt cursor and make post-resume effects idempotent.
Where the evidence leads
What this means next
Revisit a checkpointing component only after repeated verified restart-state loss or costly reconstruction makes the added dependency proportionate.