Vercel Ship 2026: a sourced recap for builders

Vercel Ship 2026 is Vercel's multi-city launch of agent infrastructure: a place where coding agents deploy software, where you run your own agents, and where Vercel's agents investigate production. It is for founders, AI architects, creators, and people who ship SaaS or an agent platform. The constraint is that the 15 October 2026 recap and the product changelogs do not always share a status label, so a launch sentence is not a license to treat a beta as finished. The primary recap is Vercel Ship 2026 recap.

This guide stays on that recap and on the product pages it points at. The runtime choices that moved on 15 October 2026, including Python, Daemons, Drives, and Connect events, are in the agent runtime guide. The working adopt, wait, and skip table is the audience decision matrix. The shorter decision essay is Vercel Ship 2026 decisions for people who ship agents.

Watch the recording

Vercel Ship 2026 is the official Vercel recording. The Ship page uses that same video as its livestream replay.

The London morning keynote is a separate recording, Keynote - Ship 26 London, and the New York opening keynote is Opening Keynote - Ship 26 NYC. Use those when you want the room. Use the recap page when you want the written claims.

What the recap says the platform is

The recap splits Vercel into three jobs. Coding agents get a deploy target. You get the primitives to build and run your own agents. Vercel uses production traffic, traces, and anomalies so its own agents can investigate on their own and then surface pull requests. The published sentence pairs that pull request with alerts: agents surface the request after they investigate. That third job is the one teams misread as "the platform will ship my product." On the London panel, Arthur Viegers of Cursor said Shopify and Amplitude already auto-review and merge 60-70% of low-risk pull requests with no developer time, while a two-line authentication change still goes to a human. That 60-70% figure is his report about those two companies. See the Ship 2026 recap.

The same page is the right scope for the tour. Published 15 October 2026, it records Ship as held in London, Berlin, New York, Sydney, and San Francisco. The opening line says thousands of people gathered across those five cities. The close says the tour returns in May for Ship 2027. The live Ship page lists speakers and announcements. Speaker names on that page are not a transcript. If a talk is not written on vercel.com, it is not a claim in this guide.

The Agent Stack

Tom Occhino's Agent Stack post is the architecture map. It says every agent needs three capabilities: connect to models, run workflows across many steps, and connect to the systems and channels people already use. The primitives on that page are:

• AI SDK is one interface for model calls, so streaming, tool calls, and structured output are not a new integration per lab.
• AI Gateway routes across models from one endpoint, fails over when a provider is down, and tracks spend. The page says you pay the provider's price with no markup.
• Workflow SDK checkpoints steps, retries the failed step, and can pause for a person, a slow API, or a webhook.
• Vercel Sandbox gives the agent an isolated microVM with its own filesystem and kernel. The page says credentials are injected when code calls a service, so the agent does not see a raw token.
• Chat SDK is the delivery layer for Slack, GitHub, and the other channels listed there, from one codebase.

Those five are the stack. eve, on the same page, is the opinionated directory layout on top of them, and that page marks eve as public beta. Do not promote eve to a generally available framework from a later memory. Re-read the Agent Stack before you standardize a team on it.

Vercel Connect, without a standing secret

At Ship, Connect was the new block: the agent asks for a temporary credential scoped to the task, instead of a long-lived provider token sitting in environment variables. That sentence is in the Ship 2026 recap and, in more detail, in the Agent Stack section on Connect.

Status moved after the tour. On 25 August 2026 Vercel marked Connect generally available on all plans. The current rule on that page is the one to implement: instead of storing long-lived provider secrets, your code requests short-lived, scoped tokens at runtime, and deployments authenticate with the existing Vercel OIDC identity. Read Vercel Connect is now generally available.

Two later Connect capabilities are still public beta, and they are not the same feature. MCP Events let an app subscribe to webhook events from a compatible MCP server. Enterprise-Managed Authorization lets an app access supported services on behalf of a signed-in user without a separate consent prompt per service. Both pages say public beta. The runtime guide treats them as wait items unless the workforce identity problem is the one you actually have.

The GA post also publishes a price: Hobby includes 500 token requests and 1,000 triggers a month at no extra cost, and Pro is billed at $3 per 1,000 token requests and $0.95 per 1,000 triggers. That is the page's price on the day this guide was written. Check the same URL before you put it in a forecast.

Containers and Dockerfile support

Malte's container announcement is Run any Dockerfile on Vercel. You add a Dockerfile.vercel. Vercel builds the image, stores it, and runs it on Fluid compute. The process must listen on $PORT. The page is explicit that the container is stateless: it takes a request, returns a response, and keeps nothing between requests. Persistent state belongs in a backing service. A Go, Rails, Spring, FastAPI, or nginx process can fit. A worker that must sit and pull jobs does not fit this shape.

That distinction matters. The October Daemons beta is the pull-based path, and it is still a public beta. Do not describe a Dockerfile.vercel HTTP server and a Daemon as one product.

Vercel Services

The recap says that with Vercel Services, the frontend and backend deploy together, with previews of the full app. The current docs page is the status check for how services call each other. Services says Services (Beta) are available on all plans. A service is an independently built unit. Public traffic reaches it only through a top-level rewrite. A binding lets one service call another without exposing that target on the internet.

Use Services when a JavaScript frontend and a Python or Go backend should share a project, a domain, and a preview. Use a single framework deploy when you do not have that split. The docs also point background workers at Daemons, which remain a separate beta. Read the Services docs before you rewrite a monorepo around the beta key.

Vercel Agent

The 15 October 2026 recap says Vercel's own agents investigate production and then surface pull requests. It does not name the dashboard product Vercel Agent or give that product a status. The changelog that matches the expanded product, published 30 June 2026, is the status check. An expanded Vercel Agent says three things are in public beta:

• Dashboard chat answers questions from the project's deployment data.
• Production investigations take a failed deployment, a runtime error, or a cost spike and recommend a fix.
• Approved actions can open a pull request, roll back, or change config only after you sign off.

The same page says Agent runs as its own identity, is bounded by the requesting user's permissions, and is read-only by default. Elevated work needs a scoped plan and an approval. Generated code runs in Vercel Sandbox before production. Availability is public beta for Pro and Enterprise, with a gradual rollout. Investigation and code-writing requests are billed as provider tokens at the provider rate, plus a Vercel token rate of $0.25 per million tokens, as written on that changelog.

There is an older changelog titled as a limited beta from 2025. It describes a different, earlier assistant. Do not cite that page as the Ship 2026 agent. The page for this behavior is the 30 June 2026 public beta note.

How to read the tour without copying it

London, Berlin, New York, Sydney, and San Francisco were customer and product sessions on top of the same primitives. The useful pattern in the recap is not a metric from a keynote slide. It is the split between a human decision and an agent action: identity on the call, a sandbox around generated code, and a preview before production. When you want the October runtime map, leave this recap and open the agent runtime guide. When you want the four calls in one table, open the audience decision matrix.