Wink Pings

The More Agents There Are, the Less They Should Hold Meetings: Vercel Board's Bulletin-Board Model

Vercel has released Board, an experimental product that uses a bulletin board to coordinate multi-agent collaboration and avoid the two pitfalls of a 'company brain' and 'group chat.' Its internal v0 software factory has already merged more than 50% of PRs, and it supports Slack/GitHub sync and fine-grained permissions.

Vercel has released an experimental primitive called Board: a secure message board for people and agents.

Multi-agent collaboration has two common pitfalls: combining all agents' memories into a shared brain, which makes division of labor meaningless; or having agents report point-to-point, where the volume of communication grows quadratically as their number increases. Board avoids both and only provides a bulletin board.

This approach is like ants. Ants do not send messages to one another; they leave information in the environment, and other ants sense it and continue the work. Biologists call this stigmergy. Board leaves tasks and progress on the bulletin board, and agents claim them themselves.

Several mechanisms are worth a closer look:

Tasks carry leases; if they are not renewed upon expiry, they are automatically returned to the pool—similar to pheromone evaporation, so a crashed agent will not hold onto work indefinitely. Labels and task dependencies act as signposts: only after a fix is completed does the review task become claimable. Watchers wake agents only when there is a new reply or a label change; otherwise they do not run. Each agent still uses its own harness and keeps its own memory, and only the traces everyone needs to see are placed on Board.

Vercel's official blog disclosed two internal use cases: a swarm of security agents uses Board to scan the codebase for vulnerabilities and push fixes; the v0 software factory has been rebuilt on Board and now merges more than 50% of PRs. Board supports two-way sync with Slack and GitHub, and permissions can be fine-grained enough that a given agent can only read and write in a specific channel; with sandboxing, real credentials do not even need to be loaded.

This model was not invented by Vercel. The open-source project cavil-loop is already doing something similar with GitHub issues + labels: tasks and discussions stay in the issue, labels determine whose turn it is, and participants are woken only when the data actually changes. When Luosky discussed this topic on X, he said something blunt: the more agents there are, the less they should hold meetings. Rather than making agents smarter, it is better to design the environment well: whoever arrives can understand the progress and continue the work.

The bottleneck in multi-agent collaboration is environment design, not model capability. The bulletin board moves context out of human brains and shared memory and into a structured environment; humans only need to make decisions when it is their turn. Some commenters noted that Board is only infrastructure—task orchestration and agent division of labor still have to be designed by users. That is true. Vercel Board has taken this step, and the direction is already clear.

发布时间: 2026-11-20 19:38