I’m building a LangChain/LangGraph-based workflow where execution is path-dependent (agent decisions determine subsequent nodes), and I’m increasingly hitting the need for a durable message/work queue between “graph events” and “worker execution.” Today, it’s possible to combine external systems (e.g., Redis/Celery/Temporal/SQS) with LangChain, but there’s no canonical pattern or first-class abstraction for: (1) routing a unit of work (node execution, tool call, LLM call) to exactly one worker, (2) persisting status/metadata for observability (queued/running/succeeded/failed/cancelled), (3) retry/backoff and dead-letter handling, and (4) resuming from checkpoints reliably without duplicating work. I’d like to propose a feature (or reference implementation) that defines a minimal “QueueAdapter” interface plus a recommended integration path (especially with LangGraph checkpointers and streaming events), so production users can plug in Redis/RabbitMQ/Kafka/Temporal with consistent semantics. Happy to contribute design notes or a PR if there’s interest.
The features you’re requesting all seem to be provided in LangSmith Deployment. It gives you a durable task queue, job statuses, support for multitasking, observability, etc. all out of the box.
What’s missing from the feature set? Would love to help out or hear more.