A platform built for agents and the apps that run them

Vercel Ship 2026, held across London, Berlin, and New York, centered on a single thesis: the next generation of software development will be driven by agents, and the platforms that support them need to be designed accordingly. CEO Guillermo Rauch framed it as deploying "software that can think," and the announcements spanned infrastructure, application development, and enterprise security.

The company’s strategy rests on agentic infrastructure with three pillars: Vercel as the deployment target for coding agents like Claude Code and Codex; Vercel as the platform for building and running your own agents in production; and Vercel itself being automated by agents that use production data to investigate anomalies and surface pull requests rather than just alerts.

Primitives for building agents

To support agent development, Vercel outlined the Agent Stack, a set of building blocks that cover the core needs of any agent: connecting to models, running multi-step workflows, and integrating with external systems and people.

  • AI SDK — a single API for any model, making streaming, tool calls, and structured output consistent across providers.
  • AI Gateway — a single endpoint for routing requests across hundreds of models, with automatic failover.
  • Workflow SDK — durable execution with built-in retries, state persistence, and observability.
  • Vercel Sandbox — isolated microVMs for agents to run and test code before production.
  • Chat SDK — deploy agents across Slack, Discord, GitHub, and other channels from one codebase.

Two new pieces round out the stack. Vercel Connect provides agents with secure, scoped access to external tools and data without storing long-lived provider tokens in environment variables. Instead, agents request temporary credentials tied to a single task, eliminating standing secrets that could leak.

Also new is eve, an open-source agent framework derived from the architecture Vercel used to build its own hundreds of production agents. eve agents live in a single directory with instructions in markdown and tools in TypeScript. Durable execution, sandboxed compute, approvals, subagents, and evals are pre-wired, so a production agent can be scaffolded, run, and deployed with familiar commands.

Expanding full-stack and backend support

Beyond agents, Vercel is broadening what can be hosted on its platform. Two long-awaited features address the backend: Dockerfile support and the Vercel Container Registry (VCR). You can now run Docker images on Vercel Functions or Sandboxes, and manage them with standard docker commands like push, pull, and tag. VCR is OCI-compliant, supports unlimited repositories per project under the same OIDC and access-token controls as the rest of Vercel, and precompiles images for fast startup.

Backend framework support has also expanded. FastAPI, Flask, Express, and Hono can now run at scale, and backend-only services such as REST APIs, durable workflows, queues, cron jobs, and MCP servers are supported. The Vercel Marketplace now includes direct access to Amazon Aurora, Aurora DSQL, DynamoDB, and OpenSearch from the dashboard.

For microservices, Vercel Services, launching July 1, makes them a first-class citizen. Frontend and backend can be developed and deployed together, with backend-only changes still producing a full preview environment. Services can now communicate with each other without traversing the public internet.

Enterprise security and governance

A significant focus at Ship was making agent and app deployment safe for large organizations. Vercel for Enterprise Apps and Agents packages the developer experience with the identity, access, and governance controls enterprises require. This includes:

  • Vercel Passport (Public Beta) — keeps internal apps private by default behind your identity provider.
  • Security Dashboard (Private Beta) — a unified view of security posture across accounts and projects, flagging misconfigurations, members without MFA, shared secrets, and long-lived credentials.
  • Bring Your Own Cloud (BYOC) on AWS (Private Beta) — run Vercel, including functions, in your own AWS tenant.

Vercel Agent for automated operations

The final major announcement was Vercel Agent, now in Public Beta. Built on eve and the Agent Stack, it acts as an intelligence layer for shipping on Vercel, monitoring production, investigating alerts and anomalies, and opening fixes as pull requests for human review.

Vercel Agent uses a permissions model distinct from typical action-by-action approvals. It plans what permissions it needs for a task and requests them in one coherent approval step. It operates under its own identity, is read-only by default, and requests narrow, temporary permissions only when it needs to touch production.

Europe gets its first Ship week

Vercel's Ship 2026 made its European debut in London, kicking off with a week of community events ahead of the main keynote. The opening stretch included a hackathon with OpenAI that drew 200 builders shipping agents on Codex and Vercel. Standout projects included Stella, which hunts unclaimed grants for London SMBs; Oscar, a co-pilot that refines prompts before they hit a model; and Phone Jail, a distraction blocker that roasts users who reach for their phones.

The schedule also packed in dinners for reporters, founders, VCs, and CTOs — with enterprise teams from the likes of Marks & Spencer and Currys in attendance — plus a 400-person Day Zero builder night featuring Mentimeter, Okta, MongoDB, DeepMind, ElevenLabs, and Cursor. A 160-person waitlist formed for CTO Malte Ubl's talk on the new software stack, followed by a conversation between CEO Guillermo Rauch and Deliveroo's Will Shu on building the agentic enterprise.

London session highlights

  • Anthropic: André Balleyguier showed Claude Managed Agents on Vercel, where Anthropic hosts the agent loop but every command runs inside a Vercel Sandbox the user owns, keeping filesystem, processes, and network egress in that environment.
  • Currys/Elkjøp: Tomas Jansson detailed a three-phase evolution of a Nordic retailer's stack from storefronts to shopping agents, anchored by a Next.js migration that cut time to first byte by 40%, aiming for a state where customer intent replaces navigation.
  • Vercel: Matan Kushner reported that Vertex, Vercel's support agent, now automates 91% of support tickets and saves 5,000 engineer-hours a month without degrading as context grows.
  • Auth0: Jas Sagoo and Sam Bellen argued that identity matters more than model quality for getting agents to production, demonstrating four standards-based patterns on the Vercel AI SDK — Token Vault, CIBA approvals, agents as first-class principals, and on-behalf-of delegation — so every action is scoped to a human decision rather than a shared API key.
  • Vercel: Abhi Sivasailam walked through d0, Vercel's internal data agent on eve, which runs analyses in a sandbox and now gets 45% of its queries from other agents rather than people.
  • OpenAI: Romain Huet showed developers how to delegate whole tasks to Codex Cloud Agents on Vercel, making the case that as agents become users of your product, you design for agent experience with scoped API keys, markdown docs, and plugins.
  • Callstack: Michał Pierzchała demoed a QA agent that runs mobile apps on real devices, explores user flows autonomously, and posts screenshots, recordings, and logs to pull requests on every PR.
  • ElevenLabs: Alex Holt paired the ElevenLabs Speech Engine with the AI SDK to build a voice agent that replies in under a second, handles mid-sentence interruptions, and switches language mid-conversation.
  • Vercel: Malavika Balachandran Tadeusz showed software shipping itself by extending the agent loop through testing, observability, and experimentation with primitives like Vercel Flags and Skills.

Agents in production panel

COO Jeanne Grosser moderated a panel with EU CTOs and AI leaders on moving agentic workloads from prototype to production. Key takeaways:

Arthur Viegers (Cursor): Autonomy should track risk — the better an agent can assess the risk of a change, the more it can run on its own. Shopify and Amplitude already auto-review and merge 60–70% of low-risk PRs with no developer time, while sensitive changes still route to humans.

André Balleyguier (Anthropic): Judge each action by reversibility and blast radius, then contain it accordingly, for instance inside a self-hosted sandbox. As autonomy increases, observability and evals become the safety nets.

Ryan den Rooijen (Currys): The predicted death of stores never happened, and Currys leads its markets by betting on humans. The open question is which shopping journeys customers want agents to handle versus a person.

Nicolas Le Pallec (AKQA): AI-native customer experience replaces navigation with intent, requiring three layers: capturing true customer intent, a brand brain storing brand and product data for AI retrieval, and generative UI that composes pages on the fly.

Fireside chat with Harry Stebbings

The day closed with a conversation between Rauch and 20VC's Harry Stebbings. From the lightning round:

On founder attributes: Stebbings said the profile has completely changed — it used to be a sales-led CEO who could fundraise and grow customers. Now he wants an engineering- and product-led founder.

On building industry voice: It's teachable but hard — a muscle built over six months of daily commitment.

On standing out when anyone can produce content: New formats, intros, and storytelling matter. Content itself has shifted; shows about leadership and hiring no longer land, whereas founder-plus-AI stories do.

Ship Berlin at the Old Mint

The Berlin event took place at the Alte Münze, where coins were once hammered by hand before machine automation — and where the JavaScript logo was born, and where Rauch first met CTO Malte Ubl at JSConf EU. Over 500 people gathered as Vercel made its case for building in the agentic era.

Berlin session highlights

  • AI SDK 7: Suyog Rao announced the release, noting the SDK is now downloaded over 16 million times a week. Version 7 turns it into a toolkit for building and running agents that reason, call tools, run across many turns, and work across files and sandboxes — all behind one provider-agnostic interface.
  • SAP: Philip Miseldine showed how the company rebuilt its design system to be composable, so people and agents assemble interfaces from the same parts, with v0 running across build, deliver, and decide stages.
  • AI SDK 7 builders: Lars Grammel, the primary engineer behind the update, focused on the new harness layer and generation beyond text. Carsten Høyer from AKQA demoed the Starbucks ChatGPT app as a "generative store" composing branded results from Starbucks' own products and identity, letting customers pull the brand into their conversation. Stephen Batifol from Black Forest Labs showed where Flux image models fit in an agentic stack: sub-second generation and editing now, heading toward real-time generation and world models.
  • Cursor: Arthur Viegers showed coding agents moving from autocomplete to working across an entire codebase independently.

CTOs on AI and growth

Chris Williams moderated a panel on building and scaling as creation costs trend toward zero:

Sven Rosemann (Flaconi): AI won't fix a broken business model — operational excellence still comes first. He likens ecommerce to a decathlon where you must be top-tier across all disciplines. Discovery is shifting into chat, so Flaconi is investing in the data foundation that makes its catalog retrievable by agents, betting buying and checkout follow.

Tjaž Silovsek (Astra AI): Venture funding is no longer a prerequisite; you can ship, find distribution, and iterate. Astra reached 10 million users and $25M ARR fully bootstrapped with a small team kept fast by a stable foundation and agents granted autonomy while kept as safe as possible.

DACH Partner Award

Vercel presented its first-ever DACH Partner Award to Valtech for exceptional impact, delivery, strategic thinking, and team enablement. The presentation also thanked sponsors Auth0, Storyblok, Cursor, SAP, Accenture Song, Blazity, and JVM.

Ship 2026: New York and Sydney

New York: The largest Ship stop yet

Ship NYC drew over 3,000 registered attendees, making it the most in-demand stop of the tour. The five-day program in New York mixed hackathons, executive receptions, workshops, and the main Ship day.

The week opened with Built in NYC, a hackathon co-hosted by Vercel and Anthropic alongside Sentry, Auth0, Resend, and Supabase. More than 100 people spent an afternoon building and demoing agents on Vercel.

Alongside the public events, Vercel hosted a series of private gatherings: a VIP reception for leaders in AI-native development, a VC dinner for founders, and a closed-door CTO/CIO dinner at Per Se where the conversation centered on where traditional architectures break down in the agent era. At the AI Social Club, Tom Occhino led a fireside chat with William Anderson (CTO, Bloomberg Media), Jeremy Bunting (VP Engineering, Serhant), and Ali Patterson (Staff Design Program Manager, DoorDash) on what's working and what isn't in agentic infrastructure for enterprises.

Ship day: Customer stories from the stage

Builders from seven companies walked through production agent deployments on the Ship day main stage:

  • Ryan Coyne (SERHANT.) presented S.MPLE, a real estate agent on Vercel that delivered a 144% increase in commission income in its first full year.

  • Brandon Bloom (Brex) gave its expense-audit agent a bash-only shell so it manages its own context instead of drowning in tool responses, cutting tool calls and token usage by 75%.

  • Ari Schapiro and Chris Sev (Auth0) demonstrated scoped, revocable agent permissions using standards-based identity patterns like Token Vault and out-of-band human approval, live on the Vercel AI SDK.

  • Jenny Wang (Alta) and Maxwell Osborne (PUBLIC SCHOOL NYC) showed how Alta's AI styling app runs on Vercel, generating an avatar per second and over 250 million outfits. They built a custom version for PUBLIC SCHOOL's New York Fashion Week relaunch.

  • Charlton Roberts (FLORA) collapsed 18 bespoke creative-production tools into two by letting an agent run and write its own code against the AI SDK on Vercel Sandboxes.

  • Prasad Wangikar (Stripe) broke down how to give agents spending power using emerging protocols including MCP, UCP, and x402 for the machine-to-machine economy.

  • Shaown Nandi (AWS) discussed Vercel and AWS working together, with native AWS databases and OpenSearch available in the Marketplace and models routed through the AI Gateway, proven by The Weather Company's 350M+ users.

Jeanne DeWitt Grosser moderated a panel on deploying internal agents in production. Andy Martin (MJH Life Sciences) argued the way to win the shadow-IT battle is to "constrain the stack and stay permissive inside it," training non-engineers on a deliberately tight stack of Claude Code, Vercel, and GitHub with ephemeral preview deployments as the guardrail: "If it becomes the Department of No, you will have people working around you." Greg Chan (SERHANT) described structuring agents to mirror a human software team — product-manager agents write requirements in Linear, engineering agents generate specs, and build-and-QA agents take it from there — with evals for hallucination and compliance, including a custom model for Fair Housing Act rules. Victor Oliveros (ZoomInfo) explained how embedding v0 in their product lets 40,000+ customers build their own workflows; agents analyze recorded calls and open pull requests so the product team decides what ships, paired with ZoomInfo's context-graph permissions so each agent only touches data its user can see.

Tom Occhino sat down with Vanessa Lee (VP of Product, Shopify) to discuss where commerce and software intersect:

Lee announced that Shopify has reduced Hydrogen to only the genuinely hard parts of commerce — analytics, optimistic cart UI, variants — and made them framework-agnostic, so Shopify's commerce primitives now drop into Next.js on Vercel via the Marketplace. She also introduced the open Catalog API, built on billions of products across millions of merchants, as "the only widespread shopping API, made for commerce," allowing anyone to embed shopping anywhere. On friction, she cited Shopify founder Tobi Lütke's early insight that the TAM was never demand-limited but tool-limited: "Lowering the friction is what releases" the untapped resourcefulness in the general population.

Closing keynote: Ivan Zhao on designing for agents

Guillermo Rauch closed Ship New York with a conversation with Notion CEO Ivan Zhao.

Zhao advised designers to design for the API first and the UI last, because "your next customer isn't only a human, it's a human and an agent, and agents are blind" — they read semantics and APIs, not interfaces. "A great tool for developers turns out to be a great tool for agents," he said. On model strategy, Zhao explained that Notion routes automatically by task intelligence, using frontier models for coding and smaller, faster ones for support and summarization, since "speed creates discontinuities" that can change the kind of interface you'd even design. Asked about a billion knowledge workers becoming agents, he cautioned that second-order effects are unpredictable: "We shape our tools, and then our tools shape us." On taste versus velocity, he chose taste: "It's the distillation of wisdom from a lot of shots on goal."

Ship 2026: Sydney

The Sydney stop opened at Machine Hall, a 1930s electrical substation that powered the city's trams for half a century. Over 200 people gathered there for Ship Sydney.

The week began with the AI SYD meetup, featuring lightning talks and networking for the local technical community. The AI Social Club brought senior leaders to Shell House for a talk from Malte Ubl and a fireside with Suyog Rao and David Hyman (Founder, monō ai) on agentic infrastructure in the enterprise.

Ship day sessions in Sydney covered regulated software delivery, voice-driven development, and organizational change:

  • Belen Garcia Fava and Vivek Katial (Heidi Health) explained how their team ships fast in regulated environments with agents and Vercel, arguing that trust is engineered, not granted.

  • Robin Yoong (Cursor) created an eve agent by voice alone — three models coded in parallel, and a spoken approval shipped it to Vercel.

  • Nonny Bell (Vercel) moderated a panel with Matthew Blode (Linktree), Aarush Issar (Auto & General Australia), and Leonardo Borges (CommBank) on adopting AI agents in production.

  • JJ Lecocq (Vercel) led a discussion on reshaping people, culture, and structure for AI.

After London, Berlin, New York, and Sydney, the Ship 2026 tour concludes in San Francisco in October. Tickets are still available.