LangGraph Production Playbook

Look forDefine a TypedDict, dataclass, or Pydantic schema. Add reducers for fields that multiple nodes write, such as message lists. Keep state small and serializable. One developer spent time fixing runtime type errors in loosely typed state. Another found schema growth the main maintenance cost.

Look forUse Postgres or MongoDB in production, because the in-memory saver loses state on restart. Qodo added Postgres persistence in a few lines. Pin the checkpointer package with LangGraph. Postgres saver 2.0.23 broke reads for one team, and another reported SSL errors under load across versions.

Look forChoose the durability mode per workload. The sync mode saves each checkpoint before the next step starts. The async mode writes while the next step runs, so a crash can lose the latest state. The exit mode saves only when the run ends. Each checkpoint adds a database write, and long threads pile up rows.

Look forCall interrupt() before irreversible actions and resume with Command(resume=...). Airbnb pauses trust cases for days until a signal or human decision arrives. One team adds a human checkpoint after each critical step. Another replaced custom checkpointing with LangGraph checkpoints and interrupts.

Look forReplaying or forking from a checkpoint re-executes every later node, including model calls and API requests. Make side-effecting nodes idempotent, with a key per action, before using time travel in production.

Look forOn LangGraph Platform, now called LangSmith Deployment, one team lost everything streamed since the last checkpoint when a user canceled a run. Test cancel and reconnect paths, and decide which streamed output you must save.

Look forCheck outputs in code and route failures back with a conditional edge. Airbnb runs a rules layer with no model in it that approves, denies, or escalates obvious cases before any LLM call. It also lets every agent abstain so uncertain cases go to a person.

Look forLangSmith traces graphs when LANGSMITH_TRACING is set, and stream modes and checkpoints feed your own tooling. One developer lost a failed production run's state before they could inspect it. Keep checkpoints for failed threads long enough to debug them.