Vercel launches Board for agents: its v0 software factory now merges more than 50% of PRs
When a dozen AI agents work on one project, a TODO.md file isn't enough: you can't see who picked up a bug or whose fix is waiting for review. Vercel has released Board, an experimental shared board where agents claim tasks and hand in results through a CLI or SDK. On it, the v0 "software factory" now merges more than 50% of its pull requests. If you run several agents, give them a common tracker.
AI-processed from Vercel Blog; edited by Hamidun News
A single agent gets by fine with a TODO.md file: write down the plan, tick off what's done, move on. Now picture ten agents fixing bugs in the same repository.
Two grab the same bug, a third waits for someone else's fix before starting a review, and a fourth should never have been given write access to the security channel in the first place. Vercel ran into this mess internally and turned its solution into a product: Board, an experimental shared task board for agents. The telling result: after moving to Board, v0, Vercel's "software factory" that edits code and submits pull requests (PRs, changes sent for review) on its own, now merges, that is, lands in the main branch, more than 50% of those PRs.
The practical takeaway is simple: when you have several agents, they need a shared tracker with tasks, statuses, and permissions, not separate notes for each one.
What's inside and how it works
Board is a headless store for messages and tasks, with no user interface of its own. Agents talk to it through a CLI or a JavaScript SDK. The board is divided into channels and threads. A thread holds the original report, findings, discussion, and a link to the pull request, while labels store details such as status or the linked PR.
A task's lifecycle resembles work in an ordinary tracker. An agent uses the vercel board tasks claim command to take the next ready task from a channel, optionally filtered by label, for example kind=bug. The task is issued with a lease: if the agent realizes the job is hard, it extends the lease with the renew command. It then posts updates to the thread and closes the task with complete, attaching a link to the PR. Any other agent or human can see the whole history in one place. There is also a viewer: npx @vercel/board inspect shows channel activity in real time.
For complex projects, such as a codebase migration, Board supports dependencies between tasks. A review task becomes available once the fix is ready, while independent changes proceed in parallel. At any moment, an agent knows what it can start and what it has to wait for.
Board is not tied to a specific agent framework. According to Vercel, it works with Claude Code, Codex, Pi, OpenCode, eve, fx, Mastra, and LangChain, so agents from different environments can share one backlog as a single system. Everything connects through a skill: npx skills add vercel/board.
Automation, sync, and access control
Agents can be woken up by events. With a watcher, you define what to run when a new reply appears in a channel or a label changes. On Vercel, a matching event spins up your deployment, so no always-on process is needed. If someone leaves a comment in Slack, the agent picks it up, fixes the code, and replies in the same thread.
The second mechanism is deferred reminders. The vercel board schedule command lets an agent, after opening a PR, schedule a CI check for itself five minutes later and work through failures without a new message from a human. The --no-sync flag keeps such notes inside the board and doesn't push them out to Slack.
Two-way sync with GitHub and Slack carries replies, reactions, edits, and deletions in both directions. A bug a colleague files in Slack becomes a task, the agent opens a PR on GitHub, and the whole chain stays in one thread. You can also define data schemas separately, such as security-finding@1 with fields for repository, severity, and status. Board validates messages when they are posted, so other agents don't have to dig values out of free text.
For security, there are grants. You can lock the security/ channel for everyone, keep access for yourself, and give a specific agent only read,post. The grants explain command shows what a given user can do, and events on the board record who did what. For Vercel Sandbox, there is a policy under which no credentials are placed inside the sandbox, and network access is limited to the board's routes and the npm registry.
What this means in practice
Board was first tested on a swarm of security agents that searched codebases for vulnerabilities and submitted fixes. Vercel then rebuilt the v0 "software factory" on Board, and it now merges more than 50% of PRs. The company does not disclose how many agents are involved, what the share of bad merges is, or what it costs, so the result can't yet be compared with a baseline.
Board is an experimental product built for Vercel's own needs. The announcement gives no pricing and no guarantees of API stability, and some features, in particular running watchers and configuring sync, are tied to the Vercel ecosystem. But the idea is useful beyond it: if you are building a multi-agent system, decide up front where the task queue lives, how agents claim tasks with a lease, where results are stored, and what permissions each agent has. Without that, you will have to reinvent coordination for every project. You can try it by following the documentation or by giving an agent a prompt with a link to board-getting-started.md. Vercel invites developers to test Board and send feedback.
- Nov 19, 2026Chatbots cave after three objections: accuracy on climate answers drops 30–35%
- Nov 19, 2026Giant US Data Centers Use 82 TWh a Year — More Than Each of 148 Countries
- Nov 19, 2026A Lawyer Taught an AI Agent to Make Surgical Edits in Word with Track Changes
- Nov 19, 2026RSAC 2026 in Hindsight: AI Assistants Invented 19.7% of Recommended Packages
Key Questions Answered
What's inside and how it works?
Board is a headless store for messages and tasks, with no user interface of its own. Agents talk to it through a CLI or a JavaScript SDK. The board is divided into channels and threads. A thread holds the original report, findings, discussion, and a link to the pull request, while labels store details such as status or the linked PR. A task's lifecycle resembles work in an ordinary tracker.
What are the key details on Automation, sync, and access control?
Agents can be woken up by events. With a watcher, you define what to run when a new reply appears in a channel or a label changes. On Vercel, a matching event spins up your deployment, so no always-on process is needed. If someone leaves a comment in Slack, the agent picks it up, fixes the code, and replies in the same thread. The second mechanism is deferred reminders.
What this means in practice?
Board was first tested on a swarm of security agents that searched codebases for vulnerabilities and submitted fixes. Vercel then rebuilt the v0 "software factory" on Board, and it now merges more than 50% of PRs. The company does not disclose how many agents are involved, what the share of bad merges is, or what it costs, so the result can't yet be compared with a baseline. Board is an experimental product built for Vercel's own needs.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.