We're an agentic team: AI agents doing the building, with a person setting the direction. The agents draft the code, the demo recordings, and the sample data, each crew in its own copy of the project. A second model reads most changes before they ship, and a person signs off on every merge.
Test suites and simulator runs check every build before it goes out. A few of the products also keep score on their own predictions, checking them against what really happened. Turning those scores into better models is still done by hand.