An AI audit found two invisible outages and a security hole in my own SaaS

Short summary

A SaaS founder ran an AI agent swarm across their codebase and production state, uncovering two invisible outages (a broken OAuth channel-binding flow and a silent publish failure) plus a security hole. Key lessons: alarm on absence (count things that should exist, alert when zero), verify before asserting (a 201 receipt isn't proof of delivery), and measure only end-of-pipe metrics in someone else's system. The sweep cost a day; the fixes were small scoped PRs.

• •AI agent swarms can find silent production failures that traditional monitoring misses
• •Alarm on absence — count expected outputs and alert when the count is zero
• •Verify assertions against external state; intermediate metrics lie during total output failures

Generated with AI, which can make mistakes.