Can You Trust a Non-Deterministic Agent to Test Your App?

Computer use agents can drive any interface a human can, from head units to mobile HMIs to companion apps, without per-platform test code. But they're non-deterministic: the same instruction can take a different path each run. The catch is that the system under test is non-deterministic too. New releases shift the UI, ops pushes environment changes, data drifts, and most test failures come from state nobody controls. So how do you trust an agent in a CI pipeline where a flaky test is worse than no test?
This talk walks through the engineering patterns that make agent-driven test automation reliable enough to gate releases. We'll cover why both the agent and the application carry non-deterministic state, how to write intent-based tests that absorb that state instead of breaking on it, and how to harness and guardrail agent runs so the output is structured, auditable evidence rather than a noisy log.