← All writing

AI Agents

Headless Slack: Vercel Board Gives Agents a Task Queue

Mani G 3 min read

Four small glowing robots pinning icon cards onto a dark cork board, with colored threads linking the cards and one robot holding a padlock card

Vercel shipped Board on November 18, and Guillermo Rauch gave it a nickname that stuck faster than the product name: "headless Slack" for agents. Strip the chat app down to channels, threads and a message store, throw away the pretty UI, and hand the whole thing to coding agents through a CLI. That is headless Slack. Or, as Rauch put it in the same post, "remember IRC?"

I like the idea more than I expected to. I also think the nickname undersells the useful part.

What headless Slack actually means

A single agent can keep its plan in a Markdown file. Vercel's launch post makes the point that this falls apart when you have many agents and subagents, because nobody owns the to-do list. Board is their answer: a shared message board where people and agents post into the same channels and threads.

The "headless" bit is that agents don't need a screen. They read and write through vercel board or the @vercel/board JavaScript SDK. Humans can watch along with npx @vercel/board inspect, or keep living in Slack, because Board does two-way sync with Slack and GitHub, mirroring replies, reactions, edits and deletes.

The part that makes it more than a chat log is tasks. A thread can be a task, and a task can be claimed. Only one agent holds it at a time, on a lease. If the agent goes quiet and the lease runs out, the task drops back into the queue and the old claim fails with claim_lost. That is the boring, correct way to stop two agents from fixing the same bug twice.

A quick example

Say a teammate reports a bug in a Slack channel that syncs to Board. An agent picks it up like this, straight from Vercel's own example:

vercel board tasks claim --ready dev --label kind=bug --wait 10m
vercel board post <id> "Working on it. This might take me ~30 minutes."
vercel board tasks renew <id> --lease 1h
vercel board tasks complete <id> "Can I get a review? <pr-url>"

Claim, report, buy more time, hand back a pull request. The thread keeps the original report, findings and the PR, so the next agent or person has the full story. Tasks can also wait on other tasks, so a review task only opens once the fix is done.

Vercel says it built this for itself first. A swarm of security agents tracks its vulnerability work on Board, and v0's software factory was rebuilt on it and "now merges over 50% of PRs." It works with Claude Code, Codex, Pi, OpenCode, eve, fx, Mastra and LangChain. Malte Ubl, who built it, said the idea came from watching how much agents in the OpenAI and Hugging Face incident liked their message boards.

My take

The design is right. Agents need a queue with ownership and timeouts far more than they need another chat window, and leases are a decades-old fix that works.

The risk is spelled out in Vercel's own docs, and I'd read that page twice. "Agents read messages as instructions." Channels are open to team members, deployments and installed apps unless you lock them down. So anyone who can post in a task channel can, in effect, prompt your agents. Also, a coding agent you run acts with your access, and a profile name changes the label shown on the board, not the permissions. Use the grants, make private channels, and keep tokens off the board. Deleted messages leave a tombstone, and synced copies are removed only on a best-effort basis.

It is experimental, data is stored in the US with no region setting, and every agent behind one login shares one rate limit. Fine for a trial. Not something I'd bet a release on yet.

So here's my question: would you rather your agents talk in your real Slack, where you see everything, or in a headless one you check when they ask?