← back
I Run a Fleet of AI Agents Across Three Machines. Here's What Broke. - Kyle Jaejun Lee, KRAFTON
Takeaway
Durable state, isolated environments, and centralized review become essential as agent fleets spread across machines.
Summary
- A three-machine fleet uses hierarchical agents with scoped contexts and approval boundaries to reduce the operator’s coordination burden.
- Missions, status, and handoffs persist in files so sessions can reset or restart after crashes without relying on context compaction.
- Failures included uncontrolled worker growth, memory exhaustion, credential collisions, laptop outages, and conflicting shared context.
- Git and SSH move work between machines, one always-on review gateway centralizes approvals, and Kubernetes is the planned foundation for scheduling and resource isolation.
agent-fleetsorchestrationpersistent-state
Original description
An honest field report from my own personal fleet of AI agents, run across several machines as a daily driver. Less about any single tool, more about the journey: how things that work on one machine break once you scale to many, what it takes to keep a setup like this running, and where it's all converging. Not a company platform — just real, evolving lessons from running it myself. Speakers: Kyle Jaejun Lee (KRAFTON): Kyle is a builder and AI Platform Engineer working to make AI agents first-class citizens in the workplace X/Twitter: https://x.com/kyleleee_119 LinkedIn: / jjlee-swe GitHub: https://github.com/cooco119