TMux

The Human Bottleneck in Agent Orchestration When we first start playing with AI agents, we run one or two in a terminal. It feels magical. But as Kyle Jaejun Lee discovered, scaling up to six active agents instantly breaks the human developer. You stop writing code. Instead, you become a glorified scheduler, an external memory bank, and a manual reviewer. To escape this trap, Lee built a strict corporate-style hierarchy: CEO, VP, manager, and worker. This is not a cute metaphor; they are concrete, code-enforced agent types. By scoping context so that it only flows down and results flow up, the human only reviews what reaches the top. You hold one context in your head instead of six. Moving Agent State Out of the Context Window Typically, an agent's state lives inside the LLM context window. When that window fills up, standard frameworks compact or summarize the history. This is slow, lossy, and unpredictable. Lee stopped compacting. Instead, he moves the state to physical files on disk within scoped workspaces. When a context window fills up in Claude, the system executes a hard reset, completely clearing the active memory. The agent then reads its own handoff files from disk and resumes work. If a machine crashes, the progress survives. Five Ways the Multi-Machine Fleet Broke Moving from a single MacBook to an active three-machine cluster (adding two headless Linux boxes) exposed severe infrastructure failures. First, orchestrator agents refused to delegate, greedily executing tasks themselves instead of spinning up workers. Lee solved this by wrapping execution in a CLI harness, forcing delegation. Second, managers opened so many worker panes in tmux that they shrunk to unreadable, un-parseable dimensions. Third, memory vanished; Claude and MCP processes devoured the MacBook's swap space. Fourth, credentials collided across shared workspaces. Finally, the MacBook occasionally went to sleep or crashed, killing in-flight jobs. Lee fixed this with file-based boot commands and an always-on Linux box hosting a unified SSH review gateway. Why the Future of Agent Compute Looks Like Kubernetes Using Discord bots as a remote control for the fleet solved immediate visibility issues, but deep architectural hurdles remain. Consistently sharing tools, managing secrets, and handling resource allocation across diverse machines are solved infrastructure problems. Instead of rewriting clustering software from scratch, Lee plans to target Kubernetes as the execution substrate. Agents should simply declare what resources and tools they need. Kubernetes handles the scheduling, secrets, and compute, while the orchestration layer sits cleanly on top.