Nov 28, Paris: AI Health Hack #2 with Alan, Qovery and AWS. Applications close Nov 21

Agentic infrastructure in 2026: where to run AI agents, their backends and databases in your own cloud

Agentic infrastructure is what AI agents run on and operate. Here is what it needs, whether Vercel BYOC is enough, and which platforms run agents, backends and databases in your own cloud account.

Romaric Philogene
CEO & Co-founder
NOV 19, 2026 · 17 MIN
Agentic infrastructure in 2026: where to run AI agents, their backends and databases in your own cloud

Key points:

  • Agentic infrastructure is the layer AI agents run on and the layer they operate. In production an agent needs long-running execution, a database, private network access to your tools, a sandbox for untrusted code, model access, and its own scoped identity with an audit trail. Most platforms cover part of that list.
  • Vercel's agentic infrastructure is the best place to build TypeScript agents inside a web app. The AI SDK, AI Gateway, Fluid compute, Workflows and Sandbox (Firecracker microVMs) fit together well. Daemons, Vercel's long-running workers, are in public beta: production only, no previews, no autoscaling, no persistent storage.
  • Vercel BYOC is not enough to run agents and their backends in your own cloud. It is Enterprise only, AWS only, and covers Vercel Functions (Node.js and Python on Lambda). TLS terminates at Vercel's CDN, requests and responses pass through Vercel, and databases, Daemons, Sandbox, containers and Services stay outside your account.
  • For agents, backends and databases in your own account, the realistic shortlist is Qovery, Northflank, or raw EKS/GKE/AKS with your own tooling. Render and Railway run agents well, but on their infrastructure.
  • Qovery runs agents, backends, databases and a preview environment per pull request in your own AWS, GCP, Azure or Scaleway account, or any Kubernetes cluster, with API Policy Tokens that scope each agent, per-agent audit attribution, Agent Skills and an MCP Server. It is not an execution sandbox, so pair it with one if you run untrusted code.

Qovery · Infrastructure control plane
Build with Claude Code, Deploy with Qovery
Learn more

In November 2026 the headline on vercel.com reads "Agentic Infrastructure". I run a company that competes for part of the same budget, so take my view with that in mind. I still think the phrase is right, and I think Vercel deserves credit for naming the category out loud.

The question I get from CTOs this month is narrower than the headline. They already decided agents are going to production. What they want to know is where those agents should run, next to which backend and which database, and inside whose cloud account. This article answers that, including the question that comes up in every second call: is Vercel BYOC enough?

What is agentic infrastructure?

Agentic infrastructure is the infrastructure AI agents run on and the infrastructure they operate. The first half is a runtime: compute that can hold a long agent loop, a database for state, queues, network access to your internal tools, and somewhere safe to execute code the agent wrote. The second half is a control surface: an API and permissions model that lets an agent deploy, roll back or scale things on your behalf without holding admin credentials.

Vercel uses the term mostly for the first half, and for coding agents shipping apps. We use it for both halves, because in production the two meet. The agent that writes a migration also runs it, and the worker that serves an agent also needs a database it can reach. I wrote a longer piece on the control-surface side in what an agentic infrastructure platform is. This one is about where the workloads live.

The stakes are real on both sides. Gartner predicts that more than 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls. In the same release it expects 33% of enterprise software applications to include agentic AI by 2028, up from less than 1% in 2024. Two of those three cancellation reasons are infrastructure decisions: what the agent costs to run, and who can prove what it did.

Kubernetes is already where most teams run AI in production. The CNCF Annual Cloud Native Survey found that 82% of container users run Kubernetes in production, and 66% of organizations hosting generative AI models use it for some or all of their inference. So the agent infrastructure question for most companies is whether agents join that estate or live on a separate platform.

What does an AI agent need in production?

A production agent needs eight things, and the gaps between platforms show up in the last four.

NeedWhy an agent needs itWhat fails without it
Long-running executionTool loops wait on models and APIs for minutes or hoursTimeouts mid-task, half-applied changes
State and a databaseConversation history, task queues, embeddings, resultsAgents that forget, or state stuck in a vendor's store
Background workers and queuesAgents triggered by events, not only HTTP requestsPolling hacks and lost jobs
Private network accessAgents call internal APIs, warehouses, ticketing, CRMsExposing internal systems to the public internet
A sandbox for untrusted codeAgents write and run code you did not reviewAgent-written code running next to production secrets
Model access and routingSeveral model providers, fallbacks, cost trackingHard-coded keys and one provider outage taking you down
Scoped identity and auditEach agent acts with its own least-privilege credential"The bot did it" in a post-mortem, with no way to tell which bot
A preview environment per changeAgent PRs touch APIs, workers and migrations, not only UIAgent changes tested against nothing, merged on hope

The first four are the classic backend list. The sandbox and model routing are where Vercel is strong. The last two, identity and full-stack previews, are where most agent projects get stuck once security reviews start, because the agent now holds credentials to real infrastructure.

What does Vercel's agentic infrastructure cover?

Vercel covers the build-and-run loop for TypeScript and Python agents very well, and it covers it on Vercel's own compute. The pieces fit together, which matters more than any single feature.

The AI SDK is an open-source TypeScript toolkit for calling models, streaming responses and defining tools. AI Gateway sits in front of many model providers behind one endpoint, so you can switch models, set fallbacks and track spend in one place. Fluid compute bills functions on Active CPU, which suits agents because time spent waiting on a model does not count. On the Pro plan, Active CPU costs $0.128 per hour and provisioned memory $0.0106 per GB-hour for functions.

For code the agent writes, Vercel Sandbox runs untrusted or agent-generated code in isolated Firecracker microVMs, with snapshots and persistent sandboxes. Drives, persistent volumes you mount into a sandbox, became generally available on October 15, 2026, at $0.05 per GB-month on Pro. For long work, Vercel Workflows pause and resume durable steps, and Queues handle message processing.

The newest piece is Daemons, long-running background processes built from container images, such as queue consumers and workflow workers. Daemons entered public beta on October 15, 2026. During the beta they run in production only, preview deployments are not supported, there is no autoscaling and no persistent storage, and you set instance counts per region by hand. They cost $0.128 per Active CPU-hour and $0.0212 per GB-hour of provisioned memory on Pro.

Plans are simple to read: Hobby is $0, Pro is $20 per month with a $20 usage credit, and Enterprise is custom. Vercel Functions run up to 800 seconds on Pro and Enterprise, with a 30-minute extended maximum in beta, and up to 4 GB of memory and 2 vCPUs (Vercel Functions limits).

If your agent is a TypeScript feature inside a Next.js product, calls hosted models, and keeps its state in a Marketplace database, this stack is hard to beat. I would not move that agent anywhere.

Is Vercel BYOC enough to run AI agents in your own cloud?

No, not for an agent that comes with a backend and a database. Vercel BYOC moves one thing into your account, Vercel Functions, and only on AWS. It does not move the data path, the workers or the storage.

Here is what Vercel's BYOC documentation says, as of November 19, 2026. BYOC is available on Enterprise plans and became generally available on AWS. It runs the Vercel Functions behind your project as AWS Lambda functions in your own AWS account, while "Vercel still routes requests, serves static assets, and manages the deployment". The Vercel CDN receives each request and terminates TLS, and your function's response returns through Vercel to the client. The supported runtimes are Node.js and Python. Functions on the Edge runtime or the container runtime, and Services, are not supported. You run the builds on your own machines or CI, and your team owns the IAM roles, the networking, the AWS quotas and the cleanup of old deployments.

Vercel is candid about the limits. Under "When not to use BYOC", the docs list the case where "all traffic must stay in your account", because requests and responses pass through Vercel, and the case where your team cannot own AWS operations.

Agent componentVercel defaultVercel BYOCIn your own account with Qovery
Agent API as a Node.js or Python functionVercelYour AWS account (Lambda)Your account, as a container
Request and response pathThrough VercelThrough Vercel (TLS ends at Vercel's CDN)Your load balancer
Long-running workerDaemons (beta) on VercelStays on VercelYour account
Agent database and vector storeMarketplace partnerMarketplace partnerManaged DB (e.g. RDS) or container DB in your account
Sandbox for untrusted codeVercel SandboxStays on VercelNot offered; pair with a sandbox provider
Containers, Helm charts, Terraform modulesContainer functions, ServicesNot supportedSupported
CloudsVercel regionsAWS onlyAWS, GCP, Azure, Scaleway, any Kubernetes cluster
PlanHobby, Pro, EnterpriseEnterprise onlyEvery plan, including Free

There is a second Enterprise option that often gets confused with BYOC. Secure Compute gives your functions a dedicated VPC with static IPs, VPC peering and site-to-site VPN. Vercel describes it as a dedicated network "that Vercel runs on your behalf". It gives your functions private connectivity to your databases, and the functions still run on Vercel.

So BYOC is enough in one specific case: your agent is a Node.js or Python function, you are on AWS, you are on Enterprise, and your security team accepts that request and response data transits Vercel. If the agent also has a worker, a database, a vector store or a GPU service that must sit in your account, BYOC leaves all of those where they were.

Run your agents, backends and databases in your own cloud.
Qovery deploys AI agents and the services around them into your own AWS, GCP, Azure or Scaleway account, or your existing Kubernetes cluster, with scoped API Policy Tokens and per-agent audit trails. The Free plan includes unlimited AI agents.

Which platforms let you run AI agents with their backends and databases in your own cloud account?

Three options run the whole agent stack inside an account you own: Qovery, Northflank, and a self-built platform on EKS, GKE or AKS. Render and Railway are good agent platforms that host for you. Vercel sits on both sides of the line depending on the plan.

PlatformWhere the agent runsBackends, workers, cronDatabases in your accountCloudsPreview env per PR (full stack)Agent permissions and auditUntrusted-code sandbox
VercelVercel; Lambda in your AWS with Enterprise BYOCFunctions, Services, Daemons (beta)No, Marketplace partnersVercel regions; AWS for BYOCFrontend and functions; Daemons excluded in betaTeam roles, deployment protectionYes, Vercel Sandbox
QoveryYour account or clusterYesYes, managed (e.g. RDS) or containerAWS, GCP, Azure, Scaleway, any clusterYes, cloned from a blueprint environmentAPI Policy Tokens (OPA), per-agent audit, read-only MCP by defaultNo
NorthflankNorthflank cloud or your account (BYOC)YesYes, in your VPC with BYOCAWS, GCP, Azure, Oracle, CoreWeaveYesTeam RBAC, API tokensYes, Kata Containers or gVisor
RenderRender's infrastructureYes, workers, cron, WorkflowsNo, Render Postgres on RenderRender regionsYesTeam rolesYes, Render Sandboxes
RailwayRailway's infrastructureYesNo, on RailwayRailway regionsYesTeam rolesNo dedicated product
Raw EKS / GKE / AKSYour accountYes, you build itYesEach provider separatelyYou build itKubernetes RBAC, you build attributionYou add one (e.g. gVisor, Kata)

Northflank is the closest peer to Qovery here, and I will say so plainly. Its BYOC mode keeps workloads, databases, logs and secrets in your VPC on AWS, GCP, Azure, Oracle Cloud or CoreWeave, and it offers sandboxed execution with Kata Containers or gVisor, plus GPU workloads. If running untrusted agent code is central to your product and you want it in your own account, Northflank covers that in one platform, and Qovery does not.

Render has leaned into agents with Render Sandboxes and Render Workflows, next to background workers, cron jobs and managed Postgres. It is a strong choice for a team that wants agents, workers and a database on one managed bill. The workloads run on Render's infrastructure, so it does not answer the own-account question.

Railway gives small teams the fastest path from a repo to a running service and database on Railway's platform. Same trade as Render: great developer experience, vendor-hosted.

Raw EKS, GKE or AKS gives you everything and asks for everything. The control plane is cheap, $0.10 per cluster per hour on EKS standard support ($0.60 per hour once a version moves to extended support). The cost is the platform you build around it: per-PR environments, RBAC that maps to your teams, a policy layer for agent credentials, upgrades every few months, and on-call. If you already have a platform team with Argo CD, Helm and Terraform in place, this is a legitimate answer.

Qovery sits on top of your cloud account or cluster and gives you the platform without building it, which is the part I know best, so I cover it in its own section.

How do you govern AI agents that deploy and operate infrastructure?

You give each agent its own credential, scope that credential to the exact actions the agent needs, record every call under that credential, and give agent changes an isolated place to run before they merge. Kubernetes alone does not do the attribution part, which is why we built it into Qovery.

Kubernetes authenticates credentials and has no idea which agent holds one. I went through the details, field by field, in how to get per-agent attribution and audit trails on Kubernetes. The short version is that if three agents share one token, your audit log shows one actor.

On Qovery, agents call the Qovery API instead of holding a kubeconfig. API Policy Tokens are organization tokens whose authorization comes from an Open Policy Agent policy written in Rego. You can allow an agent to redeploy and roll back one environment and deny everything else at the API level. Each token shows up in the audit log as its own actor, policy:<token-id>:<token-name>, so a post-mortem can tell the triage agent from the deploy agent.

The Qovery MCP Server is read-only by default. Write access needs an explicit flag in the agent configuration and write access enabled in the Qovery Console, so a curious agent cannot change production by accident. Agent Skills (qovery-deploy, qovery-preview, qovery-troubleshoot and others) give coding agents such as Claude Code, Cursor and Codex a documented way to deploy, so they do not improvise with raw cloud CLIs.

The last piece is the preview environment per pull request. When an agent opens a PR that changes an API, a worker and a migration, Qovery clones the environment, database included, runs the change in your account, and deletes it when the PR closes. I wrote about why frontend-only previews are not enough for agent-written backends in Vercel preview environments vs full-stack previews.

How does Qovery run agents, backends and databases in your own cloud?

Qovery is an Agentic DevOps platform. The control plane is ours, and the data plane is your cloud account or your Kubernetes cluster. We never host your applications.

You connect an AWS, GCP, Azure or Scaleway account, and Qovery creates and operates a Kubernetes cluster there, or you install the Qovery agent on a cluster you already run (EKS, GKE, AKS, Kapsule, k3s or on-premise, Kubernetes 1.31 and up). From there, an agent service is a container like any other. It deploys on git push, sits next to its API and workers in one environment, and gets a database from a managed cloud service such as Amazon RDS through Blueprints, or a container database for development. Secrets come from your secrets manager integration, and GPU node pools are available per service on Qovery-managed clusters when an agent needs local inference.

Because the workloads run in your account, the bill is yours too. AWS Savings Plans, Google Cloud committed use discounts and Azure reservations apply to agent compute the same way they apply to the rest of your estate.

There are three things Qovery does not do. It is not an execution sandbox, so if your agents run code nobody reviewed, keep Vercel Sandbox, Northflank, E2B or a similar isolation layer for that step. Qovery does not include a model gateway either, so your agents call model providers directly or through a gateway you choose, including Vercel's AI Gateway. Cross-cloud networking is not built in. If an agent on AWS must reach a database on GCP, you configure it with a VPN, Cilium Cluster Mesh or Tailscale.

On pricing, the Free plan covers 3 users plus unlimited AI agents, one Kubernetes cluster of your own and 1,000 deployment minutes a month, with no credit card. Paid plans (Starter, Growth, Scale and Enterprise) are on the pricing page. Bring-your-own-Kubernetes works on every plan.

Which option should you pick for your AI agents?

Pick by where the agent's data has to live and by what the agent touches, then by developer experience.

  • TypeScript agents inside a web app, hosted models, no residency requirement: Vercel. Use the AI SDK, AI Gateway and Sandbox, and keep it simple.
  • Running untrusted, agent-generated code is the product: Vercel Sandbox, or Northflank if it must run in your own account. Qovery is the wrong tool for that step.
  • Agents with workers and a database, one managed bill, no own-account requirement: Render or Railway.
  • Agents, backends, databases and per-PR environments in your own AWS, GCP, Azure or Scaleway account, with scoped agent permissions and audit: Qovery, or Northflank if its sandboxes tip the balance.
  • You already run a platform team on EKS, GKE or AKS with GitOps: raw Kubernetes, plus a policy and attribution layer for agent credentials.
  • Function code must run in your AWS account, you are on Vercel Enterprise, and traffic through Vercel is acceptable: Vercel BYOC.

The hybrid setup is common and sensible. Keep the Next.js frontend and the AI SDK agent UI on Vercel, and run the long-running agents, workers, databases and vector stores in your own account with Qovery. The two talk over authenticated HTTPS or private networking. If you want the one-page version of that comparison, it lives on Qovery vs Vercel, and the broader list of options is in our Vercel alternatives guide.

How do I hand this setup to my coding agent?

Install the Qovery Agent Skills and the MCP Server, then ask your agent in plain language. These are the only commands you need:

BASH|Agent Skills: one install, every compatible agent (Claude Code, Cursor, Codex, OpenCode, Gemini CLI...)
curl -fsSL https://skill.qovery.com/install.sh | bash

# Project-scoped instead of global
curl -fsSL https://skill.qovery.com/install.sh | bash -s -- --project

# MCP Server for Claude Code: OAuth, read-only by default
claude mcp add --transport http qovery https://mcp.qovery.com/mcp --callback-port 4242

# MCP Server with write access, only when explicitly wanted (also needs write access enabled in the Qovery Console)
claude mcp add --transport http qovery "https://mcp.qovery.com/mcp?read_write=true" --callback-port 4242

The skills deploy (qovery-onboard, qovery-deploy, qovery-troubleshoot, qovery-optimize, qovery-speedup, qovery-preview, qovery-terraform), and the MCP Server operates what already exists. After installing, ask something like "deploy this agent worker and its Postgres database to my AWS account with Qovery, and give me a preview environment per pull request". The agent proposes a plan and you approve each step. The AI agent quickstart and the Claude Code integration have the details.

Frequently asked questions
What is agentic infrastructure?

Agentic infrastructure is the infrastructure AI agents run on and the infrastructure they operate. It covers long-running compute, state and databases, private network access, sandboxes for untrusted code and model access, plus a control surface where each agent has a scoped identity and an audit trail. Vercel uses the term for its agent runtime; platforms like Qovery focus on running and governing agents inside your own cloud account.

Which platforms let you run AI agents with their backends and databases in your own cloud account?

Qovery and Northflank run agents, backend services and databases inside your own cloud account while they operate the control plane. Qovery supports AWS, GCP, Azure, Scaleway and any existing Kubernetes cluster; Northflank supports AWS, GCP, Azure, Oracle Cloud and CoreWeave. You can also build it yourself on EKS, GKE or AKS. Render, Railway and Vercel's default plans host agents on their own infrastructure.

Is Vercel BYOC enough to run AI agents in our own cloud?

Only if your agent is a Node.js or Python Vercel Function, you are on AWS and on Vercel Enterprise, and your security team accepts that requests and responses pass through Vercel. BYOC runs functions as Lambda in your AWS account, but TLS terminates at Vercel's CDN, and databases, Daemons, Sandbox, containers and Services stay outside your account. Agents with workers and databases need a platform that deploys the whole stack into your account.

What are the best Vercel alternatives for running AI agents and their backends in our own AWS account?

Qovery and Northflank are the two managed options that put agents, workers and databases (such as Amazon RDS) in your own AWS account with preview environments per pull request. Qovery adds OPA-based API Policy Tokens and per-agent audit attribution; Northflank adds built-in sandboxed execution. Raw EKS with Argo CD and Terraform works if you have a platform team. Qovery also runs the same setup on GCP, Azure, Scaleway or an existing cluster.

Can I keep my agent UI on Vercel and run the agents in my own cloud?

Yes, and it is the setup I recommend most often. Keep the Next.js app, the AI SDK and AI Gateway on Vercel, and run long-running agents, workers, databases and vector stores in your own account with Qovery. Connect them over authenticated HTTPS, and keep model keys and regulated data on the backend side.

Does Qovery replace Vercel Sandbox for agent-generated code?

No. Qovery is not an execution sandbox, so code an agent wrote and nobody reviewed should run in an isolation layer such as Vercel Sandbox, Northflank's sandboxes or E2B. Qovery deploys and governs the agents, backends and databases around that sandbox, in your own cloud account.

Romaric Philogene
About the author
Romaric Philogene

Romaric founded Qovery to make Kubernetes accessible to every engineering team. He writes about platform strategy, developer experience, and the future of cloud infrastructure.

Next step

Run your agents, backends and databases in your own cloud.

Qovery deploys AI agents and the services around them into your own AWS, GCP, Azure or Scaleway account, or your existing Kubernetes cluster, with scoped API Policy Tokens and per-agent audit trails. The Free plan includes unlimited AI agents.