World's Fair 2026
I Run a Fleet of AI Agents Across Three Machines. Here's What Broke. - Kyle Jaejun Lee, KRAFTON
Overview
This talk details the practical challenges and solutions encountered when running a fleet of AI coding agents across multiple machines. The presenter, Kyle Jaejun Lee, shares his journey from a single-machine setup to a distributed system, highlighting failures and the architectural patterns developed to overcome them. The core thesis is that managing AI agents at scale requires moving beyond a flat structure to a hierarchical organization and externalizing agent state to persistent storage, enabling robust recovery and efficient context management.
Who should watch
- AI Engineers and builders experimenting with agent orchestration.
- Product Managers looking to understand the complexities of shipping agent-based products.
- Developers facing bottlenecks with single-machine agent setups.
- Anyone interested in practical approaches to agent scalability and reliability.
Key takeaways
- A hierarchical agent structure (CEO, VP, Manager, Worker) with scoped context and approval boundaries is crucial for managing complexity, reducing human cognitive load.
- Agent state should be externalized to files on disk rather than relying solely on the model's context window, allowing for state persistence even after context resets or machine crashes.
- Implementing a review gateway where agents submit plans for explicit approval before execution prevents plan drift and provides a single control point for oversight.
- Running agents across multiple machines introduces challenges like task delegation failures, window crowding, out-of-memory errors, credential collisions, and job loss due to machine downtime.
- Solutions for multi-machine setups include using Git for context transfer, separating per-machine directories, centralizing review gateways, and leveraging Discord as a unified remote control.
- The ultimate goal is to abstract away the underlying compute infrastructure, similar to how Kubernetes manages resources, by defining agent needs and letting an orchestrator handle placement and execution.
- *The state lives in files. It is not trapped inside one model.*
- *Reuse what exists, build the new part on tops.*
Notable quotes
Kyle Jaejun Lee describes his initial setup: *I'm sitting in front of four, five, six live contexts at the same time. And here's what nobody warns you about. At that point, I'm not running agents anymore. I've become the scheduler, deciding who does what.*
On state management: *When that window fills up, the built-in move is to compact, summarize the history, make room. I stopped doing it. It's slow. I can't choose what survives.*
On the future direction: *These are the exact questions Kubernetes already answers. So, that's where I'm headed. I'm not going to reinvent compute, secrets, and tools.*
Unofficial community note. Prefer the recording for nuance.