Stop Tracking Your Agents. Give Them a Kanban Board and Watch Only the Human Queue

Instead of monitoring your AI agents, make them report their own status to a shared Kanban board, and reduce your job to checking a single "blocked by human" column. In this 82 second tip, Nate Jones lays out a lightweight management system for personal AI agents that works with Claude, Codex, or Notion. The on-screen title states the whole thesis: stop tracking your agents, make them track themselves.

Nate Jones opens with the on-screen title "AI Hack: Stop Tracking Your Agents. Make Them Track Themselves."

The Setup: One Board, Agent-Written

The mechanic is a standard Kanban board with columns you define yourself. Jones suggests to-do, blocked, doing, done, success, and blocked by human, and notes that the columns are up to you. The agent, not you, writes to the board. It uses either an MCP connection or computer use to post updates directly.

The instruction you give the agent is a standing rule, not a per-task request. Jones phrases it as: every time I work with you, write your update to the board. Morning task assignments go on the board. Task updates go on the board. The board becomes the agent's self-maintained status report.

Jones lists example columns for the board: doing, done, success

Why He Would Pick Codex

Jones says the pattern works with Claude, Codex, or Notion, but he would personally use Codex for one specific reason. Codex with computer use can both operate the computer to get into Notion directly and use Notion's MCP. That dual access matters if you run multiple Notion accounts, for example a work account and a personal account, since the agent can reach both.

This is a practical detail rather than a model preference. The requirement is an agent that can write to your board through whatever path is available, and computer use plus MCP covers more paths than MCP alone.

The Human Queue Is the Whole Interface

The key move is the blocked by human column. When an agent is stuck and needs something from you, its rule is to move that item into the human queue. That column becomes the only part of the board you have to look at.

Jones explains that the human queue is the only column you need to check

Jones pushes this one step further: you do not even have to check the queue manually. You can ask the AI to tell you what is in your human queue every couple of hours. At that point the system runs itself, and you only ever see the decisions that genuinely require a person.

Structure Is What Makes Agents Pay Off

Jones closes with the general principle behind the trick. The more structure you give the AI, the easier it is for you to keep track and actually get value out of the agents. Without a shared artifact like the board, agent work scatters across chat threads and you become the bottleneck for status. With it, tracking is a byproduct of the work itself.

Key Takeaways

Give agents a Kanban board and a standing instruction to write every update there via MCP or computer use.
Define a "blocked by human" column and make it the only thing you review.
Have the AI summarize your human queue on a schedule so you do not check the board at all.
Jones favors Codex because computer use plus Notion's MCP reaches multiple Notion accounts.
Structure is the multiplier: the more of it you give agents, the more value you extract from them.

Resources

Claude — Anthropic's AI assistant, one of the agents Jones names for this workflow.
OpenAI Codex — OpenAI's coding agent, Jones's personal pick for its computer use plus MCP access.
Notion MCP — Official docs for connecting MCP clients like Claude Code or Codex to a Notion workspace.