A controller-enforced alternative to agent-driven 'Ralph' loopsgithub.com2 ポイント·投稿者 hellointernet·6 か月前·1 コメント
hellointernet·6 か月前·議論I built this after experimenting with Ralph-style AI coding loops.Most examples let the agent own the loop and commit directly.This version keeps the loop in a controller, uses tests as the only acceptance gate, and treats git as the safety harness.Curious how others are handling rollback and verification with agents.
hellointernet·2 年前·議論I used to have a markov-chain twitter bot that was trained on a few of my friends twitter posts. It was funny because that included hastags, tagged users, and images. There were some funny results and this was long before LLMs.
Most examples let the agent own the loop and commit directly.
This version keeps the loop in a controller, uses tests as the only acceptance gate, and treats git as the safety harness.
Curious how others are handling rollback and verification with agents.