Every so often a platform announcement quietly redraws the map of what's easy to build. Vercel's Ship London 2026 was one of those. The headline wasn't a faster build step or a new framework for websites — it was that agents are now a first-class product, with a full stack behind them.
One number framed the whole event: agents now make up more than half of all deployments to Vercel's infrastructure, with the likes of DoorDash, Stripe and OpenAI building on it. The tooling has caught up with where the work is going. Here's what shipped — and why it changes how we build for the businesses we serve.
eve — "Next.js for agents"
The centrepiece was eve, an open-source framework that aims to make building an agent as straightforward as building a web app. It uses a simple file-based structure — an agent definition, plain-language instructions, plus folders for tools, skills, subagents, channels, connections, schedules and sandboxed compute.
The important bit isn't the folder layout, it's what comes built in: durable execution with crash recovery, sandboxed compute, human-in-the-loop approvals, and evals wired into CI. Those are exactly the things that separate a convincing demo from something you can safely point real customers at. Most teams bolt them on late, painfully. eve treats them as the starting point.
The Agent Stack — six building blocks
Underneath eve, Vercel formalised the Agent Stack: six components you can use together or à la carte.
- AI SDK — one unified interface to the models, with tool-calling so an agent can actually take actions, not just chat.
- AI Gateway — global routing across 100+ models, with fallover and a single place to watch spend. Swap models without rewriting anything.
- Workflow SDK — durable execution for multi-step jobs that need to survive crashes, retries and pauses for human approval.
- Sandbox — isolated microVMs so an agent's tool calls run somewhere safe, with no access to anything they shouldn't touch.
- Chat SDK — deploy the same agent across multiple surfaces — web, messaging and beyond.
- Vercel Connect — short-lived, scoped credentials instead of long-lived API keys.
Taken together, that's the difference between "we wired an LLM to an API over a weekend" and "we run agents in production with routing, durability, isolation and credentials handled properly."
Vercel Connect — the boring problem that matters most
Vercel Connect tackles the least glamorous and most important part of agents: credentials. Instead of holding long-lived API keys, agents get short-lived, scoped tokens via OIDC. The initial connectors cover the tools businesses actually live in — Slack, GitHub, Linear, Discord, Notion, Salesforce and Figma.
For anyone who has worried about what an autonomous system can reach, this is the reassuring answer: it can reach exactly what you granted, for exactly as long as you allowed, and no more.
Built for the enterprise, too
Vercel also shipped a bundle for enterprise apps and agents: Vercel Passport for identity, Enterprise Managed Users with SAML and Directory Sync, full audit trails, and Bring Your Own Cloud deployment. Governance and oversight, in other words — the things that decide whether an agent ever makes it past a security review.
The takeaway from Ship London is simple: the hard parts of running agents in production — durability, isolation, credentials, audit — are now platform features, not bespoke engineering.
Why this matters for your business
We build AI agents for service businesses — agents that answer enquiries, qualify leads, book jobs and keep records clean. The bar for that work isn't "can it hold a conversation." It's "can we trust it with real customers, real bookings and real money, with humans in control."
The Agent Stack moves that bar within reach far faster. Durable workflows mean a booking sequence survives a hiccup. Sandboxed tools mean an action can't go rogue. Scoped credentials mean we hold nothing we shouldn't. Human-in-the-loop approvals mean your team keeps the final say. We get production-grade foundations on day one — and spend our time on what's specific to your business instead of rebuilding plumbing.
That's why we've centred how we build on this stack. And it's why the agent on our site isn't a video — it's the real thing, running on it.
Talk to an agent we built on it →
The Dottie agent on our AI Agents page runs on this exact stack. Ask it about your business, or book a call and we'll map your first agent.