Agents Can Triage Your Issues. They Can't Decide What Matters.
An agent can read an incoming bug report, guess a component, assign a severity, tag it, maybe even draft a first response, all in the time it took you to read this sentence. That's genuinely useful. It's also exactly where I'd want a team to slow down and ask what they're actually automating.
Triage was never really about the mechanical act of labeling. It was about a person applying judgment:
Duplicate or related? Is this a duplicate, or does it just look like one?
Symptom or root cause? Is "the app is slow" actually a performance regression, a user on old hardware, or a symptom of the real bug in a different report?
Severity versus actual impact? Does this severity-3 report from an enterprise customer secretly matter more than the severity-1 from someone testing on their laptop?
That judgment was always inconsistent between people, which is its own problem, but at least it was inconsistent at human scale. An agent applying the same flawed taxonomy applies it to every single issue, immediately, at whatever volume comes in.
So the actual point isn't whether to use agent-based triage. It's whether the taxonomy and rules the agent is enforcing were any good to begin with.
Garbage in, garbage out was always true. It's just that "out" now happens continuously and invisibly instead of once, when a person got around to it.
What good triage discipline looks like, before you automate it
Severity definitions that are actually checkable, not "high/medium/low" left to interpretation.
Duplicate rules, written down: what makes two reports duplicates versus related-but-distinct.
An owner for ambiguous cases, so "not sure, I'll let the agent guess" isn't the default.
A regular audit of what the triage process is actually producing, not just trust that it's working because nobody's complained.
None of that is new. It's the same triage discipline good engineering orgs have always needed, the kind that's easy to skip when the team is small and everyone just knows what matters. The difference is that skipping it used to cost you some noisy backlog and occasional missed priorities. Now it costs you a system that's confidently, consistently sorting things wrong, at a speed nobody's watching closely enough to catch.
I don't think the answer is to avoid agent-based triage. I think the answer is the same as everywhere else AI tooling touches a legacy process: get the underlying discipline right first, then let automation apply it faster, not the other way around.
If your triage process has been quietly inconsistent for years and you're about to bolt automation onto it, that's worth fixing on purpose before it happens at scale.
Read: The Bottleneck Was Never Typing Speed → Get in touch →