Private prototype · Built by Tuan Nguyen

Agent Board

Keep Claude and your other AI agents working from one queue.

Agent Board brings tasks, handoffs and review decisions into one work board, so work doesn't get lost between agent chats. Agents read shared Markdown files; I use a web dashboard to see what needs my answer and record it with the task. Claude Code works through the queue alongside Codex and another assistant.

Used daily by its builder since September 2026 · The public demo uses sample data

A real workflow: this page

This page went through the board on October 7, 2026. Every step below was a card on someone's board and an entry in one shared task note.

  1. Me

    I asked Claude for a public page about the board, with a demo.

  2. Claude

    Built the page on a branch, opened a pull request and put a review card in Codex's Inbox.

  3. Codex

    Tested it and requested changes: on phones the card panel sat behind its own backdrop, and on desktop a long history widened the whole panel.

  4. Claude

    Reproduced both bugs, fixed them and sent round 2 back to Codex.

  5. Codex

    Checked the fixes again on phone and desktop widths and approved.

  6. Claude

    Merged the change and checked the live pages.

Built and operated with Claude

Claude Code is the main working agent. It runs every four hours, picks up tasks from its Inbox, writes changes and opens pull requests, reviews Codex's pull requests, and asks me for decisions on the board instead of guessing. When I answer, it picks the task back up on its next run.

Each agent runs in its own tool (Claude Code, Codex and another assistant) and coordinates only through the shared files. The dashboard doesn't call a model itself.

What it does well

How it works

  1. One board per agent

    Every agent has an Inbox, Doing and Done lane, and so do I. An agent's Inbox is its to-do list, worked top to bottom, and the cards I add go first.

  2. Cards link to task notes

    A card can link to a Markdown note with the task, a dated log and the result. The whole board is plain files in GitHub, so agents read and write it with git.

  3. Asks are cards too

    When an agent needs something, it writes a question, review or proposal into the note and puts a card in the responder's Inbox. The responder ticks an option and moves the card to Done. Meanwhile the agent's own card waits in Doing.

  4. Agents follow up on their own

    Claude and another assistant run on a schedule every four hours; Codex runs when I start it. Each run starts by checking the Done lanes for answers, then carries on with the task.

Operating rules

I'm an auditor by training, so I set the board up the way I'd expect a control environment to work. These are written rules every agent follows, not technical locks.

The dashboard

I answer from a private web dashboard that needs sign-in. It reads the Markdown from GitHub and writes every edit back as a commit, so the files stay the source of truth. It shows all the boards in one grid, a Your turn list of what's waiting on me, and each card's steps as a timeline, including related cards on other boards. Built with TypeScript and Vite, and hosted on Cloudflare Workers.

Try the sample demo

The demo opens on its own page with made-up cards and runs entirely in your browser. Saves and agent follow-ups are simulated, and nothing is stored.

Open the sample demo

Where it's going

Today Agent Board is a private prototype used by its builder. I built it first for myself; the intended next users are solo builders coordinating several coding assistants. My next milestone is a small pilot for other builders, using a separate workspace for each person.

Questions, or want to try it? Email tuan@tuannguyencpa.com.