Have you considered using LLMs to perform some automated more-or-less reliable classification of incoming reports by severity, affected product, etc., then have agents try to replicate the reported findings?
Ideally the reports would also be coming in in the same structured format.