An agent fleet across three machines: five things that broke, and an honest list of what's still unsolved
I Run a Fleet of AI Agents Across Three Machines. Here's What Broke. - Kyle Jaejun Lee, KRAFTON · Kyle Jaejun Lee
9 min total·Actually worth watching closely: ~3 min·2 must-watch clips
- 0:00 – 1:12Listen
Admitting the human is the bottleneck
He opens on the human breaking, not on tooling: with six agents live at once, you're forced to be the scheduler, the memory and the reviewer simultaneously.
What limits fleet size isn't compute, it's how many contexts one person can hold at once.
This stretch is him describing his own situation; the visuals add nothing, so headphones on a walk is enough.▶ Jump to 0:00Speaker · Kyle Jaejun Lee - 1:12 – 2:45Listen
Organizing the agents like a company
He built the agents as four real layers of entities — the top layer calls it, the middle layers break it down, the bottom executes — each with its own scoped context and approval boundary. Context flows down, results flow back up.
It isn't a company metaphor: they're real entity types, built that way so you only have to hold the single context at the very top.
He explains the hierarchy clearly and the architecture diagram just redraws what he says, so listening loses nothing.▶ Jump to 1:12Speaker · Kyle Jaejun Lee - 2:45 – 4:04Listen
State on disk, not in the model's head
Every entity gets its own workspace on disk holding the mission, the status and the handoff work product. Instead of compacting the context, he clears it completely and lets the agent read back its own handoff and resume where it left off.
A wiped context or a dead machine isn't an incident — one boot command brings the whole fleet back exactly as it was.
It's an argument about design trade-offs throughout; the screen is mostly directory structure and a few filenames you can picture yourself once you have the concept.▶ Jump to 2:45Speaker · Kyle Jaejun Lee - 4:04 – 5:25Listen
The review gateway, and the middle layer that won't delegate
Plans drift as they flow down the hierarchy, so he makes every layer submit its plan and block until approved. The counterintuitive find: orchestrator agents love doing the work themselves and would rather roll up their sleeves than dispatch.
If you can't fix the instinct, change the tools — leave only the CLIs and the skills that call them, so dispatching is the only path that works.
This is the chain of reasoning most worth hearing line by line, from symptom to cause to fix; watching the screen is more distraction than help.▶ Jump to 4:04Speaker · Kyle Jaejun Lee - 5:25 – 7:04Watch
A live post-mortem on five crashes
Managers spinning up panes until even programmatic reads came back empty, processes stacking up until memory and swap were gone, and Git credentials crossing over between workspaces.
Fleets die on isolation and resources most often; the way out is a clean, fully separated environment per workspace.
For the memory failure he puts his actual machine state on screen, and the numbers land harder than the description — worth pausing on before moving on.▶ Jump to 5:25Speaker · Kyle Jaejun Lee - 7:04 – 8:20Skim
Discord as the fleet's switchboard
Developing across machines cost him the overall mental map, so he wanted one place to root from: one Discord bot per machine, with all fleet traffic routed through it.
Once the entry point is unified, his phone becomes the remote control for the entire fleet.
The visual is a routing topology; once you've seen which machine maps to which channel, you can speed through the rest of the narration.▶ Jump to 7:04Speaker · Kyle Jaejun Lee - 8:20 – 9:10Listen
An honest ending: cross-machine orchestration isn't solved
He says plainly that single-machine orchestration works, while consistency across machines, tools stuck on one box, credential handoff and resource scheduling are all still open — the goal being that an agent declares what it needs, not where it runs.
Don't reinvent compute, secrets and tools — stack Kubernetes underneath and build only task orchestration, review flow and context management on top.
The close is judgment and trade-offs with no new visuals, but these few lines carry more weight than any slide before them.▶ Jump to 8:20Speaker · Kyle Jaejun Lee