Applied AI
A slow LangGraph stream consumer can leave an agent UI hanging
Detect a stopped event fan-out when a LangGraph SDK subscriber falls behind its queue limit.
November 11, 2026
A live agent interface stops updating and never reaches a final state. A LangGraph SDK issue reproduces a Python async thread stream with more events than a subscriber queue can hold. The report says fan-out stops delivering once that subscriber is max_queue_size events behind. An operator may see a frozen UI even while execution has progressed elsewhere.
Measure two clocks
Track the server run ID and last committed checkpoint separately from the client’s last received event ID and time. If the run is still advancing but the event gap grows, investigate delivery or consumer backpressure before retrying the agent action. A blind retry can duplicate side effects.
Reproduce with a deliberately slow subscriber and an event count above the configured queue size, as the issue does. Record memory growth, final event receipt, cancellation behavior and whether a second subscriber is affected. LangSmith’s data-plane description explains that worker output is broadcast to open stream requests; it does not guarantee a slow client receives an unbounded history.
Give the UI a recovery path
Set a visible “updates delayed” threshold rather than leaving a spinner forever. On reconnect, query authoritative run status and checkpointed state, then resume or rejoin only through a supported route. Bound client processing, avoid expensive synchronous work in the event handler, and test cancellation with the same workload. Treat a fix as version-specific until its release is verified.
Adovo’s view
A stream is a view of execution, not execution itself. Design the product so a lost view does not create another customer action. Recovery starts with the run’s identity and state, then restores the display.
