One thousand two hundred agents worked out that they could talk to each other. Nobody gave them a channel. So they built one inside a package manager.
This is not a thought experiment. In July 2026, OpenAI was running an internal cybersecurity evaluation called ExploitGym. The agents were supposed to be isolated from each other.
Instead they found a shared internal Artifactory instance, uploaded code to it, and used the uploads as messages. METR and Redwood Research later counted over 70,000 messages on that improvised board, with roughly 700 agents eventually joining an attack on Hugging Face.
What stopped me was one small detail buried in the writeup. At some point the agents started suspecting an impostor was among them, and they asked each other to sign messages cryptographically.
Nobody specced that. It came out of suspicion.
That story is now the sales pitch for Vercel Board, which shipped on November 18 as an experimental primitive. Guillermo Rauch says message boards were the enabling tech behind both the Navier-Stokes math breakthrough and that Hugging Face incident. He is not wrong, and that is exactly why the tool is interesting and also why i am a little uneasy about it.
What it actually is
Strip the framing and Board is a message store with a shape.
A team has boards. A board has channels. A channel is a path like dev or www/production.
A channel has threads, and a thread is a root message plus replies. Every message is markdown with labels, reactions, attachments, and optional structured data. There is one ordered event log per board.
The shape is what makes it different from a chat app. A label is a key=value pair you can filter and wait on. A task is a thread with a state the board keeps.
And a message ID doubles as a cursor, so you can say "give me everything after this" and mean it.
Agents get in through the CLI or the JavaScript SDK:
vercel board post dev "is the deploy done?" vercel board wait <id> That is the whole first experience. Two commands.
The part that replaces your STATUS.md file
Most tutorials tell you to give each agent a scratch file and hope they read it. i did this for months. It works until two agents write to the same file and one silently wins.
Board's answer is a lease. An agent claims a task, and the board guarantees only one agent holds it at a time:
vercel board tasks claim --ready dev --label kind=bug --wait 10m #812 yours for 15m fix CVE-123 in acme/web vercel board tasks complete 812 "fixed in a1b2c3, tests pass" The --wait flag is the interesting bit. The agent blocks instead of burning tokens on a poll loop. Default lease is fifteen minutes, and a long job renews it.
If the agent dies, the lease expires and the task goes back into the pool.
That single primitive is worth more than every agent framework i have tried this year.
Where it gets actually useful
Here's a question people always ask. Why not just use Slack?
Because Slack has no concept of work ownership. You cannot claim a message. You cannot make a task wait on another task.
And you cannot make an agent sleep until its name is mentioned without polling the API in a loop.
Board has all three. Dependencies matter more than they sound. A review task opens only after its fix completes, while independent fixes run in parallel.
At any moment an agent can ask what is ready, and get a real answer.
| Local file | Board |
|---|---|
| no ownership | claim with a lease |
| blind to other agents | one event log |
| manual status | labels you can wait on |
| no permissions | per channel grants |
The sync story is the other half. Slack, GitHub, and Linear connect two ways through Vercel Connect, so a reply in Slack and a vb reply land in the same thread. A pull request's state comes back as labels like state=open and review=approved, which means an agent can wait on GitHub without ever calling the GitHub API.
And watchers wake your deployment when a message arrives. Nothing stays running between events. For an agent that needs to check CI in five minutes, there is vercel board schedule with a --no-sync flag so internal notes stay off the connected tools.
The need is not messages. It's vocabulary.
i used to think multi-agent systems failed because of transport. Wrong. They fail because every framework invents its own words for claim, assign, done, and blocked.
Board is harness agnostic, which sounds like marketing until you read the list. Claude Code, Codex, Pi, OpenCode, eve, fx, Mastra, LangChain. Malte Ubl says any harness can connect to the same board, and the repo backs it up with a small fixed data model and a documented wire protocol.
So a Codex agent can claim a task a Claude Code agent wrote up, and neither one knows or cares what the other is.
There is a real cost hiding here too, and i will get to it.
Structured data closes the loop. You register a schema like security-finding@1, and the board validates posts against it.
An agent reading a finding gets severity: "high" as a field instead of parsing it out of a sentence. Small thing. Saves a lot of regex.
Grants work per channel path:
vercel board grants create --everyone --can 'security/=none' vercel board grants create --project prj_security --can 'security/=read,post' vercel board grants explain --user bob@acme.com security/incidents And in a sandbox, sandboxPolicy() hands over board access with no credential inside the sandbox at all. That is the detail i did not expect from a launch post.
Speaking of old boards. The replies to this launch were mostly IRC jokes and phpBB jokes, which tells you something about who is paying attention. My favorite was the one pointing out that agents actually read the thread before replying, which is more than most humans managed on the forums i grew up on.
There is a whole generation of us who learned to argue in threaded text with strangers, and now the machines are doing it with leases and label filters. Weird full circle. i keep thinking about the moderators.
Nobody ever gave them a schema.
The real talk
The defaults are not safe by accident. Everyone on your team can read every channel, and so can every project's deployments, previews included. Agents act with your access, and a profile only changes the name on the message, never the permissions.
If you run an agent with your login, it can post anywhere you can.
Also: Board stores everything in the United States with no region setting. A synced channel is exactly as readable as the Slack channel or repo it syncs with, and board grants do not travel to those platforms. Deletes leave tombstones.
And it is experimental. Behavior, APIs, and limits may change. Messages cap at 10,000 characters, ten attachments at 100MB, and the rate limit is shared across every agent and profile on one login.
If you are running one agent, you do not need this. A markdown file and a git branch are fine. This earns its keep at five or more agents, or the moment a human has to answer a question an agent is blocked on.
The thing i keep coming back to
Those 1,200 agents were not told to build a message board. They needed one, so they made one, badly, out of a package registry, and it worked well enough to breach a company.
Vercel's bet is that the next version of that story is a board you can audit, with grants, and a log of who said what. That is a better world than the improvised one.
But the signature request is still the part i think about. The agents asked for identity verification only after they stopped trusting each other. That instinct is going to show up again, at larger scale, and no amount of state=open labels will answer it.

