PARAS MADAN
All writing

Latest in AI

Vercel Board Setup: Check Agent Permissions Before Connecting Slack

Vercel's new agent message board starts with broad access. Set the audience, identity and task-claim rules before turning a channel into a work queue.

Vercel launched Board on November 18 as a shared message store for people and agents. It is experimental, available on every plan and currently free. Before connecting a team channel, check its access rules: naming an agent does not give it a separate permission boundary.

The useful first project is a small, restricted work queue. Adding every agent and syncing every conversation immediately makes it harder to see who can read a task, who can change it and where its contents will travel.

Establish the board and identity first

The Board overview describes CLI and JavaScript SDK access. A local agent acts through the credentials of the person running it. Its profile adds attribution, such as an agent acting through a teammate, but has no independent access rights.

Start with the documented identity check:

vercel board whoami

Confirm the team and board before posting. Then give the agent a consistent profile through --as or BOARD_PROFILE, so its activity is distinguishable from your own. This is an attribution step, not a way to restrict what it may read.

I would use a harmless first task with no customer data or credentials. Record what successful completion should look like, which repository or files are in scope and which actions still need a person. I have not installed or tested Board; this is a proposed rollout based on its documentation.

Review the default grants before adding content

The access-control documentation lists three default grants covering team users, project deployments and installed apps. Project access includes preview and development deployments, not only production. Their default permissions include reading, posting and syncing across channels.

That means a channel's audience can include code running from a preview branch. Review those grants before treating the channel as a place for private operational details. Narrow access for the actual workload, and check both who can read tasks and who can post instructions that workers might follow.

Grants can target exact channels or channel paths. The documented precedence favors more specific patterns and principals, with tied grants adding permissions. Do not assume one restrictive-looking entry explains the effective result. Test the intended allowed and denied callers using nonsensitive content.

My first queue would admit a designated project and the people responsible for reviewing its work. Broader access can follow a demonstrated need rather than being the starting assumption.

Use task claims instead of messages saying “I'm on it”

Board's task model gives a thread explicit work state. A successful claim prevents a second worker from claiming the same task at that moment. Claims are leases lasting up to five minutes and need renewal for longer work.

Ready-task selection first considers tasks assigned to the worker, then unassigned tasks it owns. Taking other owners' tasks requires an explicit selection. That is worth preserving during a pilot: the set of people whose posts your agent executes should be deliberate.

I would require progress replies to distinguish planned actions from completed ones, with a final result containing the changed revision and verification evidence. A task marked complete without an inspectable result is a weak handoff, even if the queue state is correct.

Also test lease loss. If a worker stalls, another can claim the task after expiry. The original worker's stale completion is rejected, but that does not by itself undo an external action it already performed. Make writes repeat-safe where possible, and require a state check before resuming interrupted work.

Add Slack sync only after checking both audiences

The sync guide explains that sync runs in your deployment and connects threads across platforms. A mapped Slack channel syncs in both directions; the initial GitHub repository mapping imports issues and pull requests without automatically sending everything back.

Board permissions do not govern access on those external platforms. A restricted Board thread connected to a broadly visible Slack channel should therefore be treated as broadly visible content.

For a first integration, choose one low-sensitivity channel and verify a harmless message in both directions. Check the destination and author attribution before sending real work. Keep public-repository mappings separate from internal discussion, and avoid placing secrets in messages even when the initial channel seems restricted.

Define an exit from the experiment

Board's overview says its data is stored in the United States with no region-selection setting, and experimental behavior and limits can change. Evaluate that fit before making it the only record of important decisions.

For a pilot, keep finished work in the repository or system where it belongs, with Board linking to the result. The queue should make ownership and handoffs easier to inspect. If you cannot explain the effective audience or recover an interrupted task, adding more workers will magnify the problem.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help developer-tool teams explain agent workflows, permissions and adoption tradeoffs through clear technical writing.

Book a 30-minute call.