Kyle Jaejun Lee develops distributed AI agent orchestration systems that keep fleets of coding agents coordinated, recoverable, and accountable to their human operators. His work addresses a practical bottleneck: adding agents often turns their supervisor into the scheduler, shared memory, and approval queue.
Lee studied at the Korea Advanced Institute of Science and Technology and held engineering leadership responsibilities at VESSL AI, where he helped introduce Hyperpocket, an open-source system for agent-tool integrations and authentication that keeps credentials within users’ infrastructure. In his announcement of the project, he demonstrated an agent authenticating with GitHub and using a Mac interface to initiate a release.
At AI Engineer World’s Fair 2026, where he was affiliated with KRAFTON, Lee described operating a MacBook and two always-on Linux machines as a three-machine coding-agent fleet. His architecture distributes responsibility across CEO, vice-president, manager, and worker agents, each with limited context and clearly assigned work. A command-line harness enforces delegation when coordinating agents attempt tasks themselves.
Three design decisions define the system:
- Disk-backed agent workspaces: Missions, status, and handoffs live in files, allowing agents to recover after crashes and rebuild their working context without relying on opaque conversation compaction.
- Centralized human approval: Agents submit plans to a single review gateway on an always-on Linux machine; approved work resumes automatically.
- Cross-machine coordination: Git, SSH, tmux, machine-specific directories, and pull requests synchronize work, while Discord bots provide a unified remote interface.
Lee’s account of operating the fleet details the failures that shaped these choices: exhausted memory, overcrowded terminal panes, crossed credentials, sleeping laptops, and conflicting machine state. His next challenge is separating an agent’s requirements from its host machine, using Kubernetes for compute scheduling and secrets beneath a dedicated orchestration, context-management, and review layer.
