Vercel Extends Its Platform for Agentic Workloads

Vercel has spent over a decade helping developers build, preview, and deploy everything from static sites to full-stack applications. That work produced the Frontend Cloud, which now supports millions of developers. At Vercel Ship 2025, the company introduced the AI Cloud, a unified platform designed for building AI features and agentic applications.

Announced alongside products built by Vercel teams working on v0 and with AI companies like Browserbase and Decagon, the AI Cloud extends the same principles as the Frontend Cloud to AI workloads. Developers get new infrastructure primitives and protections for intelligent, agent-powered applications without managing low-level cloud services.

The AI Cloud builds on the same foundation as the Frontend Cloud, extending its capabilities to support agentic workloads.

Framework-Defined Infrastructure for AI

The AI Cloud operates on the same concept that made the Frontend Cloud successful: infrastructure emerges from code rather than manual configuration. As agents and AI-assisted development accelerate code production, this abstraction becomes increasingly important. Developers and their agents build AI applications without directly touching low-level infrastructure.

At the center of this approach is the AI SDK, which standardizes LLM integration across providers, allowing developers to swap models without changing code. When deployed in a Vercel application, calls route through the AI Gateway, a global provider-agnostic layer handling API keys, provider accounts, retries, fallbacks, and performance optimizations.

An AI API endpoint built with these tools closely resembles a traditional endpoint. Standard packages accept prompts from the frontend and stream responses back. The AI SDK defines the interaction while the AI Gateway manages execution, reducing the overhead of building and scaling AI features capable of reasoning and intent detection.

import { streamText, StreamingTextResponse, tool } from 'ai';
import { z } from 'zod';
export async function POST(req: Request) {
  const { prompt } = await req.json();
  const result = await streamText({
    model: 'openai/gpt-4o', // This will access the model via AI Gateway
    prompt,
    tools: {
      weather: tool({
        description: 'Get the weather in a location',
        parameters: z.object({
          location: z.string()
        }),
        execute: async ({ location }) => {
          const res = await fetch(
            `https://api.weatherapi.com/v1/current.json?q=${location}`
          );
          const data = await res.json();
          return { location, weather: data };
        },
      }),
    },
  });
  return new StreamingTextResponse(result);
}

AI calls typically run as functions needing instant scale, but LLM interactions involve wait times and idle periods—a poor fit for traditional serverless models. The AI Cloud responds with specialized infrastructure across several layers.

Compute and Execution

Fluid Compute and Active CPU Pricing

Fluid compute optimizes for AI workloads while avoiding traditional serverless and server tradeoffs like cold starts, manual scaling, and inefficient concurrency. It deploys with the serverless model but intelligently reuses existing resources before creating new ones. With Active CPU pricing, compute rates apply only while code executes.

For workloads with high idle time—AI inference, agents, MCP servers waiting on external responses—this resource efficiency can reduce costs by up to 90% compared to traditional serverless.

Tool Execution and MCP Support

Agent reasoning is the first phase, where intent is identified and plans are generated. After that, agents execute tools. The AI SDK manages this process by registering tools, exposing them to the model, and handling execution on Fluid compute. These tool calls run either as functions in Vercel Functions or route to MCP servers.

MCP servers resemble API routes in some respects, but underneath they coordinate multiple APIs and business logic to serve as a tailored toolkit for specific AI tasks. The @vercel/mcp-adapter package simplifies this by enabling new API endpoints or transforming existing ones to serve MCP.

This initial prompt execution and reasoning run on Vercel Functions backed by Fluid compute.

Queues for Background Work

For long-running or asynchronous operations, Vercel Queues handle orchestration. Agents can fan out execution, retry failed steps, or offload background work without blocking, enabling reliable step-based workflows.

Tool calls can be thought of as functions that can be performed within the same function invocation as the agent itself, performed remotely on an MCP server, or reliably offloaded to a separate function with Vercel Queues.

Security and Observability

Running Untrusted Code with Vercel Sandbox

Agentic applications often run code generated by the agent itself, which by definition has not been validated by users. This untrusted code should not have access to production environment variables, API keys, or deployment resources.

Vercel Sandbox provides ephemeral, isolated servers for untrusted code, running on Fluid compute with Active CPU pricing. Virtual machines spin up quickly and terminate cleanly after execution, with support for Node.js and Python, pre-installed common packages, and sudo install access via a straightforward SDK.

Observation and Protection

Agentic workloads generate evolving reasoning chains that require holistic inspection. Because agentic systems loop and retry frequently, viewing single requests in isolation misses critical errors and optimization opportunities. Vercel Observability addresses this with tools for debugging slow agents, identifying hotspots, and monitoring regressions.

Agentic operations also carry direct costs through LLM provider usage—a target for sophisticated bots that execute JavaScript, solve CAPTCHAs, and navigate interfaces like real users. Traditional defenses like header checks or rate limits are insufficient. Vercel BotID provides an invisible CAPTCHA that stops browser automation before it reaches backend endpoints that trigger LLM calls or agent workflows. It works alongside the broader Bot Management suite for defenses ranging from DDoS protection to targeted application attacks.

The Shift Toward Generative Infrastructure

Web development has moved through decades of increasing dynamism, from purely static sites through dynamic applications into the current generative, agentic era. Some companies launch as AI-native from day one; others integrate AI gradually into existing applications. Both paths lead to new conversational frontends and generative backends producing content, insights, and decisions on demand.

Agentic applications follow a decades-long web transformation from purely static sites, to highly dynamic, to now agentic. We're in the early stages of this new era.

Industries from ecommerce to education to finance will build with AI, optimizing for AI crawlers and LLM SEO alongside human users. The AI Cloud represents the extension of Vercel's frontend infrastructure philosophy to this new wave of development. Developers define agents rather than just writing applications, with the platform managing the underlying complexity.