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

2026-07-08 · Source: AI Engineer · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Cloud Computing & IT Infrastructure · Depth: Intermediate, medium

Summary

Kyle Jaejun Lee, an engineer at KRAFTON, details his experience running a fleet of AI coding agents across three machines (MacBook, Linux A, Linux B) and the challenges encountered. Initially, managing multiple agents manually led to human cognitive overload. To scale, Lee implemented a hierarchical agent structure (CEO, VP, manager, worker) with scoped contexts and approval boundaries, storing agent state in disk files rather than model context windows. He also built a review gateway for plan approval. When scaling to multiple machines, issues arose, including agents failing to delegate, out-of-memory errors, credential collisions, and job loss on laptop sleep. Solutions involved a CLI harness for forced delegation, separating machine-specific state with Git for context transfer, collapsing review gateways into one central point, and using Discord as a single router for fleet control. Lee acknowledges remaining challenges like consistency and resource management, pointing towards Kubernetes for underlying infrastructure.

Key takeaway

For AI Engineers or MLOps teams scaling multi-agent systems, recognize that manual orchestration and in-model context management are critical bottlenecks. You should implement hierarchical agent structures with externalized, disk-based state to ensure persistence and scalability. Centralize review processes and leverage existing infrastructure solutions like Kubernetes for compute, secrets, and tools, rather than reinventing these foundational layers. This approach prevents cognitive overload and enables robust, distributed agent operations.

Key insights

Managing multiple AI agents requires structured orchestration to overcome human cognitive bottlenecks.

Principles

• Separate agent context for scalability.
• Store agent state externally, not in model context.
• Delegate tasks explicitly to prevent self-execution.

Method

Implement a hierarchical agent structure with disk-based workspaces for state, a central review gateway for approvals, and Git for inter-machine context transfer. Clear and reload model context instead of compacting.

In practice

• Use disk files for agent state persistence.
• Clear and reload model context instead of compacting.
• Centralize review gateways on an always-on machine.

Topics

• AI Agents
• Multi-Agent Systems
• Agent Orchestration
• Context Management
• Distributed Systems
• Kubernetes Integration

Best for: AI Engineer, MLOps Engineer, AI Architect

Related on AIssential

• [AINews] not much happened today — Frontier model constraints drive multi-model orchestration and advanced agent infrastructure for robust AI applications.
• Long-running Agents — Long-running agents overcome AI limitations by externalizing state, enabling sustained, multi-session progress on complex tasks.
• AI agents are learning on the job — just not for your whole team — Enterprise AI agents lack shared memory, hindering team productivity; shared context is vital for multi-user workflows.
• No AI Agent Orchestration Needed? What? — Modern frontier LLMs can self-orchestrate complex procedural tasks more effectively via in-context prompting than with external agent frameworks.
• Inside Anthropic’s 2026 Developer Conference — AI platforms are evolving beyond text completion to integrated, scalable agentic systems with built-in orchestration and learning.
• Weekly Dose #3 - AI’s New Bottleneck Is Control — Control over AI workflows, modalities, tools, compute, and financing is the new critical bottleneck in AI development and deployment.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by AI Engineer.