An agent can write a convincing summary of work it never finished. It can pass a quick review and then fail on the next input, because it called an API that doesn't exist or lost track of the task halfway through. "Looks good to me" won't catch that.

This session shows how to test agent behavior the way you'd test any other code. In the demo, an agent runs against a test suite with cases that should pass and cases that should fail, plus checks on what actually changed rather than what the agent says changed. A person approves anything consequential. When a check fails, we'll look at why, apply a fix, and rerun the suite to see if the fix held.

You'll leave with a way to catch agent regressions in CI and keep the final call with an engineer, so the AI-assisted work you ship is work you'd bet your pager on.