An agent's test suite is green. That is the least informative sentence in modern software engineering.
Example tests encode what you already thought of. They do not encode the inputs you missed, the data the agent changed, or the nondeterminism it introduced. Three gates catch those three failure modes before merge. I have been running this stack on agent-written patches for the past few weeks, and the ordering matters as much as the checks themselves.
Gate 1: Property checks catch the special-case agent
The first gate is the cheapest signal. You do not need a property-testing culture. You need one invariant per function the agent touched.
Here is the artifact I used on a patch where an agent rewrote parse_duration:
