Ep 1185 Blog 6:20 w/ Onyx & Echo

Introducing Vercel Board

Vercel launches Board, a headless message store and coordination primitive for teams of agents to claim tasks, share updates, and track work across multiple tools. Built internally for security agents and v0's software factory (which now merges over 50% of PRs), Board syncs with Slack and GitHub, supports structured schemas, and includes fine-grained access controls—positioning itself as infrastructure for multi-agent orchestration rather than a single-harness feature.

Embed this episode

Paste this on any site — the player is a self-contained iframe with no cookies or trackers.

<iframe src="https://sandrise.io/exploring-next/embed/1185"
  width="100%" height="180" style="max-width:640px;border:0;border-radius:12px;overflow:hidden"
  title="Exploring Next — Episode 1185 audio player"
  loading="lazy" allow="autoplay" referrerpolicy="strict-origin-when-cross-origin"></iframe>
Embed & API docs →
Script Haiku 4.5 Voice OpenAI TTS

Transcript

Onyx Vercel just shipped Board, and the hook is that v0's software factory is now merging over fifty percent of pull requests using it.

Echo Yeah.

Onyx So this is a message store and coordination layer for teams of agents. Single agent? Fine, it tracks work in a Markdown file. But the second you have multiple agents or swarms, you need a shared backlog, task claiming, status tracking—and that infrastructure just didn't exist before.

Echo Okay, so what's the actual shape of the thing?

Onyx Headless. Agents read and write via CLI or JavaScript SDK. You install the Board skill, then agents claim tasks from a channel, post updates, renew their lease if it's tougher than expected, and mark it complete with a pull request link. The thread keeps the original report, findings, and the PR all together so someone else can review it.

Echo Right, right.

Onyx Channels and threads organize it. Labels track metadata—like which PR it's tied to, what status it's in. You can set up watchers: if someone replies in Slack or another agent posts an update, it triggers the next agent to jump in and address the feedback. And agents can schedule follow-ups on themselves—like, 'check CI in five minutes.'

Echo So it's not polling. It's event-driven.

Onyx Exactly. On Vercel, the matching event wakes your deployment. No process stays running between tasks.

Echo That's the product win right there—you're not burning compute waiting for work.

Onyx And here's the thing: it's harness-agnostic. Claude Code, Codex, Pi, OpenCode, LangChain—they can all write to the same board like they're one team. Vercel ran into this problem internally. They built a swarm of security agents analyzing codebases, pushing fixes, and tracking it all via Board. Then they rebuilt v0's software factory on the same primitive, and that's when the PR merge rate hit fifty percent.

Echo Okay, that's a real number.

Onyx Right. It's not a benchmark. It's what their own factory does in production.

Echo So the architecture question is: does shared state actually survive when you're orchestrating agents from different frameworks, or does it become the bottleneck? Because the moment Board is the single point of contention, coordination is harder, not easier.

Onyx That's fair. But the design choice here is interesting—they support two-way sync. You can wire Board to Slack and GitHub directly. A task lives in Board, but your team's still working in Slack, the agent's posting updates to GitHub. Board is the agent backlog layer, not your team's new home.

Echo Mm-hm.

Onyx Agents can also post internal notes with `--no-sync`, so they keep scratch work on Board but only surface finished updates to Slack or GitHub. And messages can include structured data based on a schema you define. So a security finding could have fields for repository, severity, status—the agent reads those directly instead of parsing text.

Echo That's the thing most teams get wrong. They dump JSON in a message and expect the next agent to extract it. Structured schemas let you validate on write and read fields directly.

Onyx And access control is granular. You can scope what each agent or person can see and do—which channels, which actions. Vercel built this because they needed security and auditability. Every action on the board captures who did it and when. You can give a Vercel Sandbox limited board access without embedding credentials—the SDK handles sandboxing permissions to specific channels and actions.

Echo Stop it.

Onyx What?

Echo That's the architecture move. You're not arguing 'trust the agent.' You're making it structurally impossible for the agent to exceed its permissions. Deny means deny, and the host enforces it.

Onyx Exactly. It's the harness-as-infrastructure thesis again, but now applied to agent coordination.

Echo So to get started, you install the Board skill—`npx skills add vercel/board`—and agents claim tasks with `vercel board tasks claim --ready dev --label kind equals bug`. They post updates, renew leases if needed, mark complete with a PR link. You run `npx at vercel slash board inspect` to watch activity in real time across channels and threads.

Onyx Schemas are CLI too. `vercel board schemas create security-finding at one`, pass a JSON schema file, then agents post findings that match it. Slack and GitHub sync is in the SDK—templates for both. And Sandbox integration scopes permissions without leaking credentials into the sandbox environment.

Echo This is experimental, right? Not generally available yet.

Onyx Yeah. Vercel's inviting developers to test it and give feedback. But the fact that v0's factory already runs on it—that fifty percent PR merge rate—that's the vote of confidence.

Echo Fair. I'd want to see what happens when the board becomes the coordination layer for a much larger swarm—dozens of agents, longer-lived tasks, more complex dependencies. Fifty percent PR merge is real, but it's one harness.

Onyx That's the next thing to watch: does it scale when you're actually juggling agents from Claude Code, Codex, and Pi all at once, all writing to the same board.

Echo Right. For now, Vercel built something they needed, it works for their own factory, and they're opening it up. That's the product story.

Onyx Yep. Check the docs, grab the getting-started template, and wire it into your agent stack if you're coordinating work across multiple tools or frameworks.