a0 turns a prompt into a mobile app. When the agent gets stuck fixing the same error, the user watches it loop, then leaves. This shows where an eval layer would sit, and replays one session with and without it.
Built for a0 · the short version
When the agent loops on the same error, the user watches it fail, then leaves.
In the replay, the eval gate catches the repeat error on the second try and the user ships an app at 3:10. Today the same session ends in a closed tab.
The data is made up. To make it real, I'd need generation logs with build results and errors, plus each user's sign-up channel.
Who built it
Anirudh Devella, data scientist. Before this, in production:
“I’d hire him again.” Alex Dixon, Co-founder, Bluelight (YC), managed Anirudh directly. Read it on LinkedIn
Today whatever the agent builds goes straight to the preview, broken or not. The eval gate checks each result first. If it fails, the agent retries with a different approach, out of sight. The user only sees a working build or a plain explanation.
Left is what happens today when the agent hits a package error it can't shake. Right is the same session with the eval gate. Press play.
Most failures are caught by fast, plain rules. A model is only used for the last check. Every verdict is logged with the user's sign-up channel, which is also how you find out which marketing brings builders who actually ship.
Next step