Do I need to rewrite my whole app to use AI SDK 6 or Workflows?

No. This isn’t a framework migration. You can drop the SDK into an existing Node/Next.js project and start adding agent logic piece by piece. Think “add one endpoint,” not “burn your repo.”

What’s the difference between calling an LLM directly and using Vercel’s Gateway?

Direct calls = 12 different SDKs, random rate limit errors, and billing spreadsheets that make finance cry. Gateway = one endpoint, unified billing, better limits, and failover when your favorite model decides to take a nap.

Is Workflow basically just cron jobs with marketing?

If cron jobs could pause, wait for humans, resume a week later, retry intelligently, store state, survive redeploys, and not explode during outages — sure, call it cron. Otherwise, no. This is orchestration without the headache.

Can I build my own agent instead of installing one?

Yep. AI SDK + Workflows + Gateway + Sandboxes = everything you need to build a custom agent. Think of it as LEGO blocks, except nothing breaks when someone bumps the table.

Ship AI, AI, AI

Right after the Next.js 2025 conference, Vercel’s founder and CEO, Guillermo Rauch, walked back on stage for another conference, because apparently, one was not enough. Next.js 16 gave us faster builds, smoother navigation, and fewer “why is this cached from last week?” incidents. Cool release, solid polish.

But Ship AI was the real plot twist.

“We’re moving from pages to agents.”

Nice headline, but what actually grabbed our attention was the AI Gateway. This means there is no need to refactor the entire backend because Anthropic has launched a better model, or juggle random SDKs and pray that your finance team doesn’t see the token bill. The Gateway is the difference between “we’re experimenting” and “we’re shipping this in production without setting the office on fire.”

Durable workflows, or the newly created acronym by them: AaaS. Not what you are thinking, it stands for AI agents as a service. And why Vercel wants an “AI agent on every desk.” But before we get into this, here’s a tiny recap of Next.js 16, because yes, all of this happened back-to-back, and the internet barely survived.

Next.js conf recap

Before Vercel told everyone to stop building pages and start shipping agents, they quietly dropped Next.js 16. Think of it as the calm before the Ship AI storm. Faster builds, smoother navigation, and smarter caching. Basically, Vercel cleaned the kitchen before setting it on fire the next day.

If you missed our full breakdown, we covered all of it here: Read the blog

Turboback is the default,
Smoother navigation,
Caching is now opt-in,
React 19.2 has cleaner effects and UI,
DevTools now speaks MCP,
Build adapters API to tailor how your app builds.

Ship AI Vercel conf takeaways

We finally get to the main event. Vercel talked. We listened. And now we’re going to write about it so your team doesn’t have to pretend they “totally watched the stream.” Or forget all this and call us and tell us to create what you want.

At first, this whole “AI in production” thing felt quite simple. Drop in a model API, send a prompt, get some text back, and tell your CTO you’ve “shipped AI.” But this lasted only 48 hours. Suddenly, you’re switching models every month, juggling each provider’s weird SDK, managing rate limits, reconciling billing systems, and explaining to finance why you blew $800 on tokens because your intern forgot a .stop() condition.

With Vercel's AI SDK, you can stop duct-taping your backend together with ChatGPT responses. It’s already hitting ~4M downloads per week, just to let you know, which means either everyone is building agents or everyone is equally confused and panicking together. Either way, good news.

From here, every other announcement makes sense: workflows, durability, sandboxed execution, Gateway, and an actual path to put “an AI agent on every desk.”

AI SDK 6 & Gateway

With one SDK, you get all the models. You can switch Claude → GPT → Groq → random stealth model Vercel got early access to, by changing like three lines of code. Basically, it is the polar opposite of how you're managing models right now. Before Gateway, every time a new model dropped, you rewrote the code, like, for instance, Anthropic is better? Rewrite. GPT is suddenly smarter? Rewrite again. Gemini releases something insane next month? Refactor everything one more time. At the end, your “AI integration” starts to look like a graveyard of half-removed SDKs and unused wrappers.

The Gateway gives you:

unified billing
better rate limits
failover when your favorite model has a meltdown
and observability.

However, please note that AI SDK 6 is still in beta, and we are currently experimenting with Gateways to build agents. We will definitely share our findings on how it performs in real-world cases.

Workflows

Ship AI 2025 had one core message: "Your one-prompt agents are cute, but the real world needs grown-ups." And we totally agree with them. Agents' functionality is no longer limited to responding. They have to:

call tools
fetch data
wait for humans
retry on failure
pause for a week
wake back up like nothing happened

Yes, everything Neo failed at.

Anyway, this is where every team quietly cries, because building that has always meant a spaghetti bowl of queues, cron jobs, retries, broken state, and a prayer you don’t lose tokens mid-execution. So Vercel introduced the Workflow DevKit, enabling developers to build long-running, reliable AI agents without manually wiring tons of infrastructure.

You write one TypeScript function, mark steps inside it, and Vercel turns it into a durable workflow. This allows agents to:

Call models and tools
Respond to webhooks or events
Wait for human approval
Resume exactly where they left off

Vercel Sandboxes

LLM models can generate code, and the codes are used to deploy programs. But sometimes it can break catastrophically. The Sandbox is created to solve it, and here is how:

Every execution runs inside an isolated micro-VM.
No access to prod environments.
No “oops we just deleted our own database” moment.
Perfect for validating agent-generated fixes before a human approves them.

AI Cloud

This is where the entire foundation of Vercel's vision of adding AI Agents to every desk is designed with infrastructure to back it up. And we have already tried this for one of our clients, and we are absolutely loving the results.

This does not mean that Vercel is asking you to build agents from scratch; you can if you want to. But what it means is that you don't need to create different infrastructure.

Their AI Cloud is designed to remove infrastructure from the equation.

AI Gateway: With this, you can pick any LLM model and switch it anytime you want.
Fluid Compute: Only pay when CPU is actually used.
Sandboxes: Execute untrusted code safely.
Workflows: Durable and resumable implementations.
Observability: See every step, every token, and every failure.
Zero config: Deploy and forget.

AI Agents for enterprises

The real problem for enterprises isn’t whether we can build one, but where we even start. Which model? Which tools? How do you run this in production without duct-taping Lambdas, cron jobs, and random queues together?

This is exactly where Vercel’s “AI agent on every desk” vision starts to make sense. Enterprises don't need to hire a research lab just to send an email seven days later or validate a document. Vercel’s answer is to give companies two paths, that is either install agents, or build them, but without building infrastructure.

First, there’s Agent-as-a-Service or AaaS, which is exactly what it sounds like (Not the acronym). Instead of writing orchestration logic, retries, state management, or approval flows, you can install agents that already run natively on the AI Cloud. Vercel even launched an Agent Marketplace, where teams can grab production-ready agents from companies like Code Rabbit, Sorcery, Corid, Browser Use, Dcope, Kernel, Cubix, Mixat, and others. They’ve even built their own internal agents for code review, anomaly investigation, and debugging in production.

We also implemented a website anomaly investigation agent for one of our customers, and you can see the results below. We are damn proud of this.

And if none of those solve your exact problem, you can still build your own. Or connect with us, and we will help you get started.

Final thoughts

The whole point of their “AI agent on every desk” plan and this release is that enterprises shouldn’t need to learn distributed systems just to automate repetitive work. You either install an agent or ship your own on top of infrastructure that already works.

And yes, we follow the same blueprint internally, because if a platform can run week-long workflows, data lookups, approvals, retries, and sandboxed execution without blowing up billing, we don’t see a reason to reinvent that.