Stacks and Flows

Architecture Boundary Guide · Source-verified

Temporal with LangGraph: which durability problem belongs where?

LangGraph can checkpoint graph state; Temporal can durably coordinate execution and retry activities. Combining them may help when both boundaries matter, but it does not make external writes exactly once.

Execution boundary

Separate reasoning state from durable execution

Request
start durable run
Temporal workflow
invoke reasoning
LangGraph reasoning
commit through activity
Activity + side effect
recover from history
Persist / retry / resume
LangGraph structures agent reasoning and state. Temporal owns durable progress, retries, timers, and recovery around the work.

What the sources establish

  • LangGraph documents checkpoint-based persistence for graph state.
  • Temporal documents event-history replay and activity execution as separate parts of its model.
  • Temporal announced an integration path for LangGraph workloads.

Two forms of state

Graph checkpoints represent application progress. Temporal event history reconstructs workflow execution. They are related, not interchangeable.

Failure boundary

A replayable workflow can resume orchestration after worker failure. A database, payment, or message side effect still needs its own idempotency strategy.

Do not add both by default

The integration is justified when the workload has both a graph-state problem and a durable coordination problem. Otherwise it adds another runtime to own.

Decision detail

  1. Start with LangGraph persistence when resumable graph state is the primary requirement.
  2. Add a durable execution system when long-running coordination, failure recovery, and operational ownership justify it.
  3. Make external writes idempotent because activities may retry.

Evidence

These claims are source-verified. We do not label this page hands-on or benchmarked because no reproducible test artifact is attached.

Continue the decision