Multi-agent systems fail less on reasoning than on orchestration, says one production team

njanChe1 · reddit · 2026-07-23

The hard part of multi-agent systems, the author argues, is not the agents themselves but failure handling when one worker dies mid-task.

The post says most tutorials skip the real problem: workers OOM’ing, planners emitting impossible tasks, and races between results.
What worked in production was treating orchestration as a distributed-systems problem:
message bus plus durable queues
typed task contracts
an aggregator that waits on a pre-registered task set
stateless, single-purpose workers that never call each other
If a worker dies, its task simply remains on the queue.
The key split is clear: the model decides what to do, while durable infrastructure guarantees it actually gets done.

The author asks what others are using for orchestration: custom code, LangGraph, Temporal, or something else.

More from coding & agent