What is Vercel Sandbox?
Sandbox is Vercel’s name for isolated virtual-machine environments where agents and applications execute code, distinct from the Serverless Functions that handle ordinary request/response web traffic. Vercel’s homepage markets the broader category as “Agentic Infrastructure” and explicitly promotes systems that “reason, execute code in isolation, run for hours, and recover from failure.”
That positioning is verifiable. The specific capacity is not. Vercel does not surface per-sandbox CPU, memory, or pricing ceilings on its public marketing surface, and the homepage lists only “Sandboxed Environments” as a feature without attaching resource limits or tier pricing. Any concrete sizing figures circulating outside Vercel’s own pages are unconfirmed.
The strategic intent does come through. Vercel is spending real engineering surface on isolated execution. Its GitHub organization hosts Eve, “The Framework for Building Agents,” alongside the AI SDK and a Workflow SDK for durable, observable agents. A company maintaining a durable-agent framework next to an isolated-execution product is telling you where it expects billable compute to land.
How do Vercel’s agent-facing products fit together?
Vercel’s agent-facing surface spans model access, execution, and orchestration, per the company’s homepage and GitHub organization. The components as listed:
- AI SDK, described on Vercel’s GitHub as “The AI Toolkit for TypeScript,” the tooling layer for building AI-powered applications and agents.
- AI Model Gateway, listed on the homepage as the model-access feature.
- Sandboxed Environments, the isolated execution surface.
- Workflow SDK, Vercel’s GitHub “Workflow SDK: Build durable, reliable, and observable apps and AI Agents in TypeScript,” the orchestration layer.
- Eve, Vercel’s GitHub “The Framework for Building Agents.”
The useful detail is that these are separate components rather than one monolith. A team can adopt the AI SDK without using Sandboxed Environments, or run Sandboxed Environments without Eve. The homepage pitches “Tools to connect your agents to Vercel infrastructure,” with a $npx plugins add vercel/vercel-plugin workflow, as the connective tissue. Whether that integration is tighter than bolting the same components together on raw cloud compute is a claim Vercel makes implicitly but does not benchmark.
How does Sandbox differ from Serverless Functions?
Serverless Functions handle short-lived request/response traffic and scale to zero between invocations. Sandboxed Environments are built for the opposite shape of workload: long-running, isolated execution that outlives a single HTTP request. Vercel’s homepage frames the capability as systems that “execute code in isolation, run for hours, and recover from failure,” a description serverless functions cannot honestly meet.
This matters because agent workloads are a poor fit for the serverless billing and lifecycle model. An agent loop holds a working set in memory, calls tools, waits on sub-tasks, and persists state across turns. A model server keeps weights resident. A batch evaluation job churns through a dataset for tens of minutes. None of these want a cold start on every invocation, and none want to be billed per-request as if they were a webhook handler. Run an agent loop on a function with a 60-second timeout and the platform will kill it mid-tool-call; that is not a bug, it is the contract.
The second-order consequence is that any platform serious about agents has to offer something that looks like an always-on container, which is the gap Sandboxed Environments fills. Vercel’s own “Durable Orchestration” and “Fluid Compute” line items, both on vercel.com, reinforce that the company is building for stateful, long-lived processes rather than only for stateless edge functions. The product names do the telling here: you do not brand a tier “Fluid Compute” if your target workload is a 50-millisecond redirect.
What does it cost, and what is missing?
Pricing transparency for sustained, high-memory agent workloads is the gap Vercel has not closed. Vercel does not publish per-vCPU or per-GB-hour rates for Sandboxed Environments on its public pages. A builder trying to compare Vercel against AWS Fargate or Lambda on a cost-per-agent-hour basis will not find the numbers on vercel.com.
The funding picture is verifiable. Wikipedia records a $300 million Series F in September 2025 at a $9.3 billion valuation. That is not a pricing datum, but it explains why a frontend-hosting company can credibly stand up isolated execution: it has the money and the stated intent to expand past edge delivery.
The missing piece, and the one most likely to determine adoption, is a direct cost comparison against the alternatives teams already run. AWS Lambda and Fargate, Fly.io, Modal, and Render all host long-running agent workloads today. Vercel’s pitch is developer experience and integration with its existing deploy pipeline, and that integration is genuinely valuable when your app and your agent share a codebase and a CI path. Whether it is worth a premium, and how large a premium, is precisely the question Vercel’s public pages do not answer. Developer-experience premiums are real and defensible, but they are defensible only against a published number, and Vercel has not published the number.
When do static containers make sense for agent workloads?
Static, always-on containers earn their keep when an agent’s working set, uptime requirement, or model-server footprint makes per-invocation serverless billing uneconomic. The decision rule is straightforward: if your workload spends most of its wall-clock time holding state rather than answering requests, a container is the right shape, and serverless functions are the wrong one.
Vercel’s published customer evidence is consistent with this pattern. Vercel’s homepage states that Notion “powers millions of agent conversations daily on Vercel,” and that Mintlify powers documentation for 20,000-plus companies on the platform. Conversational agents at Notion’s volume are exactly the long-lived, stateful workload that container-grade isolation exists to serve. The Zapier figure, over 100 million monthly visits per vercel.com, is closer to a traffic story than an agent-runtime story, but it does establish that Vercel is willing to quote raw scale when it suits the pitch.
The honest thesis is narrower than a spec-sheet comparison. Vercel is building an agent runtime around isolated, durable execution, and the Eve framework plus the Sandboxed Environments feature are the evidence. Per-sandbox capacity and pricing are not published, so any concrete sizing figures circulating outside Vercel’s own pages should be treated as unconfirmed. The signal is real. The specifications are pending verification against Vercel’s own documentation.
Frequently Asked Questions
Does Vercel Sandbox require a Next.js codebase?
No. The July 2025 NuxtLabs acquisition shows Vercel expanding framework support beyond Next.js. Teams running Vue, Svelte, or raw TypeScript can adopt Sandboxed Environments independently from their frontend framework.
How do I compare Vercel Sandbox pricing against AWS Fargate or Lambda?
You cannot directly compare from public documentation. Vercel does not publish per-vCPU or per-GB-hour rates for Sandboxed Environments, while AWS publishes on-demand, Spot, and Savings Plans pricing transparently. Any cost comparison requires a Vercel sales quote or a real billing invoice from a deployed agent workload.
What happens if an agent exceeds its sandbox capacity?
Vercel does not document OOM behavior, vCPU throttling, or graceful degradation for Sandboxed Environments. The only published behavior is the general promise that agents can recover from failure, not what triggers the failure in the first place. Teams running model servers or large in-memory agents must treat capacity limits as unknown until they hit them in production.
Is the Eve framework required to use Sandboxed Environments?
No. The components are unbundled. You can run Sandboxed Environments with the AI SDK only, with a custom orchestrator, or with vanilla TypeScript. Eve is optional tooling, not a gatekeeper for the execution tier.
What signals that Vercel is serious about sustained agent workloads beyond edge functions?
The September 2025 Series F round raised $300 million at a $9.3 billion valuation. That capital funds infrastructure expansion. Vercel could have stayed in edge delivery only. Instead, it maintains an agent framework, a Workflow SDK, and a Sandbox product alongside its compute investments.
