Matthew Boston

Use AI agents to break your app

March 21, 2026

Key takeaway: Point AI agents at a staging copy of your app to find the bugs no one would think to script.

Your test suite tells you the app works as designed. To find out whether it holds up once it’s deployed, point AI agents at it.

Most suites walk the happy path, with maybe one sad path for an invalid email. Real users double-click, go back, paste garbage into fields, and use your app in orders you never imagined. A scripted test can’t model that, because someone has to think of each scenario first. If you’d thought of it, you’d probably have a test already.

An agent can roam the app semi-randomly, chaining flows and trying odd inputs. It isn’t fuzzing. Fuzzing throws random bytes at an interface. An agent knows what a form is, what a button does, and what a valid-but-unusual input looks like, so it makes traffic that looks like real users behaving unpredictably.

That finds the “impossible” bugs: the race that only fires when two actions land in the same 200ms window, the state corruption that takes an exact sequence of navigation and submits, the memory leak that shows up after one pattern of page changes. Those are the bugs users find in production on a Friday afternoon.

To make it practical, you need:

  • A realistic environment where agents can roam without touching production data
  • Good observability, so you can trace what the agent did when something broke
  • Structured chaos, so agents explore broadly but still do meaningful things
  • Automated triage to separate real bugs from expected failures

This doesn’t replace your suite. It’s a layer on top. Unit tests check correctness and integration tests check contracts. And attackers are already pointing agents at other people’s systems, so pointing them at your own first is the cheaper option.