Harvey engineer Joey Wang reports that scaling internal AI agents broke backend infrastructure rather than the models. Managing massive code volumes required building fair-share job queues, automated repair loops, and strict guardrails to handle skyrocketing pull requests.