Mateo, this is exactly the comment I was hoping the post would provoke. You're right that I framed the piece around state explosion as the headline risk — observability at the transition level is the natural next layer, and arguably the one that bites teams later and harder.
The item I'd underline from your list is the "completes but drifts" failure mode. It's insidious precisely because there's no exception to catch and no oracle for the correct path: a green run looks identical whether the routing was right or quietly wrong. The teams that handle it well stop treating routing decisions as ephemeral control flow and start persisting them as data — the decision, the inputs that drove it, and the confidence behind it, captured at each conditional edge. Once that history exists, "why did it take this path three weeks into runtime" becomes a query instead of an archaeology project.
Maker/checker divergence rate is my favorite leading indicator there. When checker disagreement starts trending up before any incident fires, that's usually drift announcing itself early.
I'm actually drafting a follow-up that goes deep on this — state snapshots, edge-traversal metrics, and divergence monitoring as first-class telemetry. Mind if I cite this comment as the prompt for it?
One production challenge I rarely see discussed in LangGraph conversations is graph observability at the state-transition level.
State explosion is definitely a real issue, but we've also seen teams struggle with understanding why a graph reached a particular decision path after weeks of runtime evolution. Once you have conditional routing, retries, human review loops, and multiple LLM-powered nodes, debugging becomes less about individual node failures and more about reconstructing state history.
A pattern that has worked well is treating state transitions as first-class telemetry:
What's interesting is that many production incidents aren't caused by crashes at all. The graph executes successfully, but gradually starts taking unexpected paths because classification confidence, retrieval quality, or model behavior drifts over time.
In that sense, monitoring graph behavior becomes just as important as monitoring infrastructure. A pipeline that completes is not necessarily a pipeline that remains correct.