← tackworks.net

Why I Built a Task Board for AI Agents

I manage a team of 13 AI agents. They handle everything from customer interactions to health checks to memory management. They run 24/7 on a mix of local hardware and cloud inference. And for the first few weeks, coordinating them was a disaster.

The problem wasn't the agents themselves. The models are capable. The problem was everything around them: who's working on what, what's blocked, what needs a human decision before it can move forward.

Human tools don't fit

I tried using existing tools. Trello, Linear, GitHub Projects. They all assume the user is a human with a browser. The API is an afterthought bolted onto a visual interface.

AI agents have different needs:

• Token budgets. Every API response costs tokens. A full Trello card with comments, attachments, and activity history burns context window for no reason. Agents need sparse-field reads — ask for id,title,status, get exactly that.
• Atomic coordination. Two agents grabbing the same task is worse than no agents at all. You need atomic claiming with TTL expiry — not a "assigned_to" field that agents politely check.
• Structured decisions. When an agent hits a decision point, it needs to present options to a human in a way the human can click and the agent can parse. Not a Slack message that gets buried. Structured options, clickable buttons, machine-readable response.
• No SDK. Agent frameworks change every month. The tool spec should be a markdown file the agent reads, not a pip package that pins you to a framework version.

What I built

Tack is a Kanban board built for this reality. One Python file, SQLite database, FastAPI server. Give your agent the TOOL.md file as context and it knows how to use the board — create cards, move them, claim them, add trace notes, request decisions.

The key features are the ones that don't exist in human tools:

• Sparse-field reads: GET /api/cards?fields=id,title,column_name returns 5x less data than a full card fetch. Your agent polls the board every 10 minutes without burning its context window.
• Atomic claiming: POST /api/cards/{id}/claim returns 409 if another agent already has it. No race conditions, no duplicate work.
• Human-in-the-loop decisions: Agent posts structured options. Human sees clickable buttons. Click one, agent gets the answer as JSON. No natural language parsing.
• Plan approvals: Agent proposes a plan, human approves or denies. Denied cards move to blocked with the reason logged. Simpler than decisions — just yes/no.
• Batch operations: Create three cards, move two, add a note — one API call. Agents that need to do multiple things per turn don't need multiple round trips.

The philosophy

If your agent needs an SDK to use your tool, your tool is too complicated.

Every Tackworks product follows the same principles:

• One file. server.py is the entire application. No project structure to navigate, no imports to trace, no build step.
• SQLite. No database server to install, no connection strings to configure. The data is a file on disk.
• TOOL.md. The agent integration spec is a markdown file. Paste it into your agent's context. Done. Works with OpenAI, Anthropic, Ollama, llama.cpp, or anything that reads text.
• Self-hosted. Your data stays on your machine. No cloud account, no API key, no usage limits.

What's next

Tack is the task board. Chock is the approval queue — for when "approved" isn't nuanced enough and you need contingent conditions. Spur is the event relay — when something happens on the board, the right people (and agents) hear about it via Telegram, Slack, Discord, or webhooks.

Together they form a lightweight coordination layer for AI agent teams. No vendor lock-in. No cloud dependency. MIT licensed.

If you're building with agents and struggling with coordination, try them. If they don't fit your needs, open an issue and tell me what's missing.