I Run a Fleet of AI Agents Across Three Machines. Here's What Broke. - Kyle Jaejun Lee, KRAFTON

Scaling AI coding agents requires moving beyond simple terminal interactions to a structured, hierarchical organization. Managing multiple agents manually creates a bottleneck where human attention becomes the limiting factor. By implementing a CEO-to-worker hierarchy, agents operate within scoped contexts and approval boundaries, significantly reducing cognitive load. State persistence is achieved by moving data out of model context windows and into file-based workspaces, allowing for system resets without losing progress. Hardware failures and resource constraints necessitate a distributed approach, where tasks are offloaded to headless Linux machines and controlled via a centralized gateway. Ultimately, the transition from manual management to a robust, automated infrastructure—leveraging concepts like Kubernetes for scheduling and resource allocation—is essential for maintaining reliable, scalable agent operations across diverse computing environments.