What Vercel shipped in 2026: the Agent Stack, containers, and a platform that stopped being a frontend host

Vercel Ship ran across five cities in 2026 — San Francisco, New York, London, Berlin and Sydney — and the announcements landed steadily enough that the shape only becomes obvious in aggregate. Taken together, this was the year Vercel stopped being describable as a frontend host.

Here is what shipped, and what the pattern underneath it looks like.

The Agent Stack

The organising idea of Ship 2026 was the Agent Stack — Vercel's collection of AI tooling assembled into something coherent rather than a pile of SDKs. It comprises the AI SDK, AI Gateway, Vercel Sandbox, the Workflow SDK and the Chat SDK.

The piece that makes it a stack rather than a bundle is Vercel Connect, which replaces long-lived credentials with scoped, short-lived tokens and full audit trails. An agent that needs to touch Slack or GitHub requests a temporary, task-scoped credential instead of holding a standing secret.

This deserves more attention than it got. The unsolved problem in production agents is not reasoning quality — it is that giving an autonomous process a permanent API token with broad scope is an unreviewable security decision. Every agent demo quietly assumes a .env file full of god-mode credentials. Short-lived scoped tokens with an audit trail is the boring infrastructure answer, and boring infrastructure answers are what move a category from demo to production.

eve, announced as generally available, is the framework that implements the stack: an open-source, filesystem-first runtime where an agent is a directory of markdown instructions and TypeScript tools, with durable execution, sandboxed compute, approvals and evals built in.

AI SDK 7 rounds it out, reframing the SDK as a toolkit for agents that reason, call tools, run across many turns, and work across files and sandboxes — rather than a library for streaming chat completions. Vercel cited over 16 million weekly downloads.

Containers, and what that admission means

The announcement that says the most about Vercel's direction is the least glamorous: Dockerfile support and a Vercel Container Registry. The registry is OCI-compliant and works with ordinary docker push, pull and tag. Images are optimised for Fluid compute and stored as precompiled snapshots so Functions and Sandboxes start quickly.

For a platform whose original pitch was that you should never have to think about containers, shipping a container registry is a real strategic concession — and the right one. The gap between "deploy your Next.js app" and "run our production workload" was always the stuff that does not fit a function signature: a Python service with system dependencies, a binary someone compiled in 2019, a browser automation job.

The same theme runs through expanded backend framework support — FastAPI, Flask, Express and Hono running at scale, including REST APIs, workflows, queues, cron jobs and MCP servers.

Vercel Services

Vercel Services, available since July 1, 2026, makes microservices first-class: multiple backends and frontends deploy as part of a single project, communicate privately without touching the public internet, and a backend-only change builds a full preview environment.

That last detail is the one that matters in practice. Preview deployments were Vercel's best idea, and their historic limit was that they covered the frontend while the backend stayed a shared staging environment everybody fought over. A preview that spans services is a genuinely different review experience.

Enterprise, and the AWS question

The enterprise announcements point somewhere specific:

• Vercel Passport — identity provider integration and private-by-default internal apps (public beta)
• Security Dashboard — account and project security posture monitoring (private beta)
• Bring Your Own Cloud — run Vercel inside your own AWS tenant (private beta)

BYOC is the notable one. Running Vercel in your own AWS account is aimed squarely at the objection that ends enterprise deals — data residency, existing committed spend, a security team that will not approve another vendor holding production traffic. It also changes the reliability calculus: your Vercel deployment's fate becomes coupled to your own AWS account's health rather than only to Vercel's.

The marketplace also gained direct access to Amazon Aurora, Aurora DSQL, DynamoDB and OpenSearch, which continues the shift away from Vercel operating its own database products and toward brokering the ones people already use.

Vercel Agent

Vercel Agent — in public beta — monitors production deployments, investigates anomalies and opens pull requests for review, using a permissions model that pairs planning with approval.

The approval gate is the design decision worth noting. An agent that can investigate an incident and propose a fix, but must have the fix reviewed by a human, is a considerably more defensible product than one with write access to production. Whether it earns trust depends entirely on its false-positive rate, which no announcement can tell you.

Shopify and Hydrogen

Vercel and Shopify redesigned Hydrogen, isolating the genuinely hard commerce problems — analytics, optimistic cart UI, product variants — into reusable primitives that work with any framework, with Shopify's Catalog API enabling product search and commerce embedding anywhere.

Framework-agnostic commerce primitives are more useful than another storefront starter. The hard parts of commerce are hard regardless of what renders the page.

August kept going

The changelog since Ship has not slowed:

• Sandbox Managed Images (Aug 10) — versioned, open-source base images with Node.js, Python and coding agents preinstalled, now the default runtime
• Bun.serve as a Functions entrypoint (Aug 10), including native WebSocket handlers
• Grok Imagine Image 2.0 on AI Gateway in preview (Aug 8), with improved text rendering and layout planning
• Public container registry repositories (Aug 7), read-only for Vercel account holders
• Audit log drains to Datadog, Splunk and Panther for enterprise teams (Aug 7)
• Skill Packs via the skills.sh directory (Aug 7), for curating and updating agent skills in one command

The pattern

Read in sequence, 2026 was Vercel answering the question "what happens when the thing you deploy is not a website". Agents need durable execution, sandboxes, scoped credentials and audit trails. Backends need containers, private networking and real framework support. Enterprises need to run it in their own cloud.

The competitive consequence is that Vercel's comparison set has changed. Against Netlify the argument was about frontend developer experience. Against a container platform it is a different and harder argument — but it is the one Vercel has now chosen to have.

The reliability consequence is worth stating plainly: a platform that runs your frontend, your backend services, your agents and your sandboxed compute is a much larger single point of failure than one that served your static assets. Consolidation buys real velocity, and it concentrates risk. Both are true.

You can watch Vercel's live status here alongside AWS — which, for anyone on BYOC, is now part of the same dependency chain.

---

Sources:

• Vercel Ship 2026 recap
• Vercel changelog