Vercel's Agent Stack: the building blocks for production agents
June 17, 2026 · AI Automators
Every agent — whether it answers support tickets or writes code — needs the same three things: a way to reach models, a way to run multi-step work reliably, and a way to connect to the systems and people it serves. Build all of that yourself and you're choosing between vendor lock-in, glue code, or your own leaky abstractions.
Vercel's Agent Stack is its answer: a set of composable primitives for each of those layers. Here's what's in it and where each piece fits.
Connect to models
Agents don't run on one model. Cost, latency, and capability trade-offs differ per task, so you need one interface to reach any model and a way to route between them.
The AI SDK is that interface — framework- and model-agnostic, covering text, images, speech, and video. Switching from one model to another is a one-string change in your code, not a rewrite, so you can move a task between Claude, OpenAI, or anything else without touching the rest of the agent.
AI Gateway is the routing layer — a single endpoint across hundreds of models that fails over when a provider goes down and tracks cost and usage as it goes. You pay the provider's price with no markup and can bring your own keys. Real-estate firm SERHANT. runs three models from one key: analysis to Claude, copy to GPT, images to Gemini.Execute complex workflows
Agents run sequentially, sometimes for hours, often executing code along the way. Two primitives cover durability and isolation.
The Workflow SDK checkpoints every step, persists state, retries failures, and pauses to wait on a person or a slow API. When a step fails deep in a run, you resume from the last good step instead of re-running — and re-paying for — every model call before it. Creative-tools company FLORA fans a single session across 50-plus image models this way.
Vercel Sandbox gives each agent its own microVM — a full Linux box with a filesystem, Docker, and its own kernel, isolated from the host and from other sandboxes. Credentials are injected only when the agent's code actually calls a service, so the agent never sees a raw token. That separation matters most when the code being run was itself written by a model.Connect to data and tools
An agent that only talks to models can't do much. It needs scoped access to your systems and a way to reach users where they already are.
Vercel Connect replaces the long-lived, do-anything token with short-lived tokens minted per task and scoped to the permissions you grant. Every action traces from user to agent to service in the audit log. It's in public beta with Slack, GitHub, Snowflake, Salesforce, Notion, and Linear — plus anything else reachable over OAuth or an API. Chat SDK delivers your agent across Slack, GitHub, Linear, WhatsApp, and Discord from one install, handling each platform's adapter for you. NanoClaw runs one agent across a dozen-plus channels, keeping context as a conversation moves from Slack to GitHub to Linear.eve: the stack as a framework
After building hundreds of agents, Vercel decided agents have a recognizable shape and turned it into eve — an opinionated open-source framework that wires the whole stack into a single directory. An agent becomes a handful of files: the model it runs on, an instructions file for who it is, and folders for the tools it can use, the skills it knows, the subagents it delegates to, the channels it lives in, and the schedules it acts on by itself. Durable execution, sandboxed compute, approvals, and delivery are already plumbed in underneath, so you write the agent and little else.
The takeaway
The Agent Stack is Vercel betting that "production agent" becomes a standard build target with standard primitives — the way "web app" did. Whether you adopt the whole stack or pick a single layer, the components are modular and most are usable today. The harder question is rarely the primitives; it's knowing which jobs are worth handing to an agent in the first place. If you want help putting any of this to work, browse our directory of providers to find people who build it for a living.