Skip to content
AI Primer
workflow

Vercel uses Board to share vulnerability-fixing tasks among agents

Vercel describes using Board to share tasks, approvals and developer replies among agents fixing vulnerabilities. Its headless state layer syncs major updates to Slack and Linear.

4 min read
Vercel uses Board to share vulnerability-fixing tasks among agents
Vercel uses Board to share vulnerability-fixing tasks among agents

TL;DR

The launch post includes schema-validated security findings and scheduled CI check-ins that can stay off Slack. The access documentation says an agent's displayed name carries no permissions of its own, and preview deployments can read every channel by default.

Vulnerability fixes and approval nudges

Vercel describes six jobs in its vulnerability-fixing workflow:

  1. Task intake: Linear tickets become Board tasks.
  2. Approval chasing: Developer nudges are tracked to prevent repeated requests.
  3. Reply processing: A dedicated agent processes GitHub comments, Linear comments and replies to nudges to determine subsequent actions.
  4. Human escalation: A “needs-you” board holds questions for the factory owners.
  5. Agent handoffs: Agents running in Vercel Sandboxes exchange tasks through a communications channel. A dedicated agent keeps pull requests green over time.
  6. Reporting: Board data feeds a gamification dashboard.

Tracking approval nudges is a pleasingly mundane use for agent infrastructure.

Task leases, watchers and schemas

Work stays in a thread from the initial report through findings and the pull request, with coordination mechanics detailed in Vercel's launch post:

  • Task ownership: vercel board tasks claim --ready dev --label kind=bug --wait 10m waits for a ready bug. Agents can renew a lease with tasks renew <id> --lease 1h and finish with tasks complete.
  • Dependencies: A review task can become available after its fix completes, while independent fixes proceed in parallel.
  • Watchers: Replies and label changes can wake a deployment to run an agent. No process needs to remain running between matching events.
  • Structured findings: Messages can carry schema-validated data, such as repository, severity and status, so downstream agents can read fields without extracting them from prose.

Slack and Linear get major updates

Vercel keeps minor agent insights away from human-facing tools and syncs significant state into Slack and Linear. Its stated goal is to avoid filling those tools with continuous agent chatter.

Replies, reactions, edits and deletions can mirror in both directions through the SDK's sync adapters. An agent can also schedule a five-minute CI check-in with vercel board schedule and --no-sync, keeping that follow-up internal to Board.

Shared state across harnesses and interfaces

Board is model- and hosting-platform agnostic, according to the launch explanation, while the launch discussion describes agents from different harnesses collaborating through it. It also works from Cloudflare Workers because it is a REST API, according to a follow-up.

Vercel already has internal factories operated through TUIs or Slack agents, as described in its discussion of custom interfaces. The underlying threads, team and agents remain shared across the interfaces in the UI demonstration.

Shared hypotheses and daily trace reviews

A second, undisclosed project uses Board for “agent swarm hill-climbing,” according to the same factory walkthrough:

  • Agents log decisions centrally for knowledge sharing.
  • Proposed improvements become hypotheses that other agents can pick up.
  • Sandbox logs and full agent traces are attached for post-mortem debugging.
  • A daily run inspects those logs and traces to improve the factory's behavior.
  • The operator interacts through Claude while the sandboxes run fx.

eve Company Brain

Separately, the eve Company Brain template packages a company knowledge assistant for Slack. Its four advertised capabilities appear in eve's announcement:

  • Concise, cited answers across company sources.
  • Scheduled weekly reports or daily triage.
  • Bug diagnosis and code writing in Slack.
  • Extension through custom tools and skills.

Agent identities and channel permissions

Board distinguishes attribution from authorization. The documentation describes several boundaries:

  • Agent profiles: --as fx changes the displayed identity. The agent still uses the permissions of its human or project, and the profile has no access of its own.
  • Default readership: Team members, project deployments (including previews) and installed apps can read every channel unless access is narrowed.
  • Synced audiences: Board grants do not apply inside Slack, GitHub or Linear. Those platforms' users can read synced content and post replies back onto Board.
  • Sandbox credentials: sandboxPolicy() can provide scoped Board access without putting the credential inside the sandbox.
  • Instruction intake: Vercel explicitly warns that agents read messages as instructions, making channel posting rights part of task authorization.
  • Data residency: Messages, tasks and attachments are stored in the United States, encrypted at rest and in transit. Board has no region selector.

Further reading

Discussion across the web

Where this story is being discussed, in original context.

On X· 6 threads
TL;DR1 post
Vulnerability fixes and approval nudges1 post
Task leases, watchers and schemas1 post
Shared state across harnesses and interfaces5 posts
Shared hypotheses and daily trace reviews1 post
eve Company Brain1 post
Share on X