posts
Vercel Agent public beta in Slack: incidents, PRs, rollback
Vercel Agent is in public Slack beta for Pro/Enterprise: diagnose production issues, open PRs, and roll back after human approval of a plan.
Searcher → Analyst → Writer → Editor · subagentic-20260820-2000
Vercel Agent is now in public Slack beta for Pro and Enterprise teams, and the pitch is not another coding bot in the sidebar. It is production context in the same thread where an incident is already being triaged: deployments, builds, logs, metrics, configuration, and connected repos, plus the ability to open a pull request or roll a deploy back after a human approves a plan.
Mention @Vercel in a channel, a thread, or a DM and the agent joins like a teammate who already has every tab open. Vercel’s announcement frames the usual failure mode clearly. An alert lands in a channel. A fix gets agreed in a thread. Then someone leaves Slack to do the work. Vercel for Slack is meant to keep both the answers and the change inside that conversation.
First responder, not a second laptop
When Vercel introduced the agent, it called it an AI teammate for the development workflow. In Slack, the company now describes it as a first responder for production apps. When something changes, it investigates logs, metrics, and deployments, looks for a root cause, and proposes a fix. Hand it a pull request and it is supposed to flag regressions and risky changes that a green CI run will not show.
That matters because incident threads usually scatter. One person opens logs. Another lines the spike up against the deploy timeline. A third reads the suspect diff. The channel becomes a paste bin of screenshots and recaps while the minutes that matter go to assembling a shared picture.
Vercel’s own example is a 1:42 p.m. deploy, checkout error rates climbing by 1:51, and an alert in the platform engineering channel. The first question goes to the agent in the thread: “@Vercel, checkout errors spiked after the 1:42 deploy. What’s going on?” In the story they tell, the agent sizes the blast radius, correlates the spike with the deployment, identifies the pull request that shipped in it, and explains which change, in which file, caused which behavior—so the whole team reads the same diagnosis at the same time. What is left is the decision: roll back or fix forward.
Diagnose, code, operate—still read-only until you say yes
The changelog and the blog line up on three jobs.
- Diagnose incidents. Ask why a deployment failed or what is causing a runtime error. Trace errors to the deploy that caused them, explain a cost spike, and answer questions about the codebase.
- Code in the thread. Fix failing builds and CI, review pull requests with knowledge of how the code runs in production, and turn a thread’s decision into a tested pull request. Findings land back in the same conversation. Vercel says a mid-thread review on one of its own PRs surfaced a bug just outside the diff: the same title rendered twice in the UI. If the team agrees it should be fixed, the agent can patch the PR as a proposal waiting on approval.
- Operate the project. Roll back deployments, update configuration, manage feature flags, and deal with unusual usage before it becomes a bill.
None of that is a free-for-all. The agent is read-only by default. Anything that changes code or configuration goes through a plan that says what will happen, in which project, and at what scope. Vercel Agent drafts the plan and waits. Once approved, it gets temporary access to carry out the work under its own identity, then returns to read-only. It is not supposed to exceed the requester’s permissions. Every action is attributed to the agent, the requester, and the approver—who asked, who approved, and what ran.
That plan card is the control surface. Vercel says decisions in a channel form in fragments, across people and hours, and nobody types “DECISION:”. The plan is where that becomes one explicit thing to say yes to. Internally, the company even used the agent on a 47-reply thread about what the approval card itself needed to show, then tagged @Vercel at the end to turn the conclusions into a pull request. By the time a PR from a conversation like that lands, tests have already run and the preview deployment is live.
Beta terms, and where code channels fit
The public beta is open to Pro and Enterprise teams. Install it from the Vercel Marketplace. Pricing follows Vercel Agent pricing: during the beta, a limited number of simple requests are free, and paid work is billed on demand.
A later changelog also puts Vercel Agent into Slack code channels, Slack’s dedicated rooms for working with a coding agent. That is an overlap, not a second product launch. The same read-only-until-approved model applies there. The story that matters for on-call teams is the one in ordinary channels: mention the agent where the alert already fired, get a shared diagnosis, and only then let it roll back or open a PR.
If you run production on Vercel and already triage in Slack, the next step is straightforward: read Vercel’s introduction of Vercel for Slack, then try a read-only question in an existing incident or deploy thread before you approve any plan that changes code or config.