Agents upload a short recording of what they built, humans leave timestamped feedback, and the agent can retrieve that feedback and iterate.
Basically, instead of an agent telling you "done", it can show and you have a nice way to share the resulting videos in PR, on Slack, with your teammates etc.
Very happy to hear what you think of it. I've built it specifically for reviewing outcomes instead of impementations.
Humans define what correct looks like, agents implement and attach evidence (image, videos, etc.) that humans can review.