agents · ai · building

The real agent problem isn't writing one, it's coordinating many

Vercel just shipped Board, a shared message board for agents and people. It's aimed at a problem anyone running more than one agent runs into fast: they can't talk to each other.

Robert MillNovember 20, 2026 · 3 min read

Abstract illustration of connected nodes representing agents coordinating on a shared board

Vercel shipped something this week called Board: a shared message board that agents and people both post to. My first reaction reading the announcement was that this is a real developer problem right now — so many agents running at once, and so little coordinating them in both directions.

A single agent is easy to reason about. It has a task, it does the task, maybe it writes its progress to a markdown file so you can check in on it later. That covers the solo case fine. The trouble starts the moment you have more than one agent, or more than one agent plus a person, all touching the same piece of work.

Where it actually breaks

Vercel's own writeup is specific about where they hit this: a swarm of security agents scanning their codebases for vulnerabilities and pushing fixes, and v0's software factory, which now merges over half its pull requests without a person writing the code. Once you're running agents at that scale, you need answers to boring questions that get hard fast. Who's working on this task right now? Did anyone already try this fix and have it fail? Is a human supposed to review this before it ships, or did we already decide that's not needed for this kind of change?

A single agent's todo list doesn't answer any of that. You need something closer to a shared inbox: threads, labels, who can read what, a record of who did what and when.

Why a message board, not a database

What's notable about Board is the shape they picked: channels and threads, the same mental model as Slack, rather than a database schema. Agents claim a task, post an update, renew a lease if it's taking longer than expected, and complete it with a link to the result. Anything in the thread, the original report, what was tried, the pull request, stays attached to it for whoever looks next, human or agent.

That's a deliberate choice. A database is more structured but less legible. Nobody wants to query a table to find out why an agent did something. A thread you can just read. And because it syncs with Slack and GitHub both ways, the record doesn't live in one more silo — a teammate can report a bug in Slack, an agent picks it up on Board, and the pull request shows up where the team already looks.

The part that matters more than the feature

The interesting bit isn't the message board itself. It's the admission behind it: coordination between agents is now its own layer of infrastructure, not something you bolt on later. Vercel built Board for their own internal swarms before they shipped it as a product, which is usually a good sign — it means the primitive got shaped by a real problem instead of a guess at one.

It also fits something I keep noticing as more of these systems ship: the gap isn't model capability anymore, it's everything around the model. Access control for who can touch what. A record good enough to audit. A way for an agent to wake back up and check on something it started five minutes ago without a human having to remember to ask it. None of that is glamorous, and all of it is the actual work of making multi-agent systems trustworthy enough to leave running.

If you're only building with one agent, you won't feel this yet. The moment you add a second one working alongside it, or a person who needs to see what it's doing, you'll want exactly this: a shared place to look, not a guess at what happened in someone else's terminal.