What Is an AI-Native Cloud? The Hosting Slice Agents Can Call

Search for AI cloud and you get a messy SERP: GPU marketplaces, “AI platforms,” and classic hyperscaler landing pages. For builders of agents, the useful definition is narrower — and it is a slice of AI agent infrastructure, not the whole stack.

An AI-native cloud (also called agent-native cloud) is cloud infrastructure that an agent can provision and operate as a first-class API consumer — apps, data stores, auth, functions, compute, queues — under the same governed identity, metering, and Account Kit policy as every other real-world action.

Naïve’s product surface for that idea: Cloud Infrastructure for AI Agents and cloud primitives.

AI cloud vs AI-native cloud

If the operator is still only a human in a console, you have cloud near AI. If the operator can be an agent calling one API without long-lived AWS keys in context, you have an agent-shaped cloud surface.

What has to be true

1. Provision is an API — create app, database, queue in code (naive.forUser(id).…).
2. Metering matches agent economics — work billed on the same credit balance as other primitives.
3. Identity is shared — the same tenant user that sends email or holds a vault entry owns the cloud resources.
4. Governance is outside the prompt — Account Kits and the governance gateway enforce allow/deny, budgets, and approvals.
5. No scavenger hunt of provider keys — the platform holds upstream credentials; the agent holds a capability handle.

What Naïve ships in this lane

From the cloud infrastructure solution and cloud batch:

• Apps — fullstack apps on a real URL (introducing apps)
• Database / storage / auth / functions — backend pieces agents need to ship (introducing backend)
• Compute + queue — Docker services, jobs, schedules, and async work (compute, queue)

This is an opinionated wrap, not a general hyperscaler console for agents. Positioning is deliberate: wrap commodity hosting; own the regulated agent-profile bundle and governance. Under the hood there may still be serious compute vendors; the product boundary is agent-shaped.

Approvals and failure modes

Giving an agent a raw cloud IAM role fails in familiar ways: keys leak into logs and prompts, idle always-on resources dominate the bill, audit trails split across vendors, and revoke means chasing roles by hand.

On Naïve, treat cloud like any other dangerous capability:

• Enable only the cloud primitives the tier needs on the Account Kit.
• Expect compute.create and compute.exec to return pending_approval (HTTP 202) for agent callers on real (non-default) tenant users by default — same pattern as cards.create and connections.connect.
• Out-of-kit calls return forbidden; unconnected third-party tools are a Connections problem, not a cloud one.

A minimal path you can run today

You need a workspace key (nv_sk_...) from Studio → Settings → API keys. Enable cloud primitives on the customer’s Account Kit first.

import { Naive, isPendingApproval } from "@usenaive-sdk/server";
 
const naive = new Naive({ apiKey: process.env.NAIVE_API_KEY! });
const agent = naive.forUser(customerId); // real tenant user, not the workspace default
 
// Ship one app under the customer's governed identity
const app = await agent.apps.create({
  name: "acme-ops",
  type: "fullstack",
});
 
// Compute is approval-gated for agent callers by default
const created = await agent.compute.create({
  name: "nightly-job",
  type: "job",
  image: "python:3.12-slim",
  command: ["python", "-c", "print('ok')"],
});
 
if (isPendingApproval(created)) {
  // HTTP 202 — surface created.approval_id; approve() replays the frozen create
  console.log("pending compute:", created.approval_id);
}

Sanitized 202 body when compute.create is frozen (SDK PendingApproval shape — ids redacted):

{
  "status": "pending_approval",
  "approval_id": "apr_01HY…",
  "action": "compute.create",
  "primitive": "compute",
  "title": "Create compute workload “nightly-job”",
  "message": "This action requires human approval before it can run."
}

Keep spend and approval guardrails tight before widening blast radius. Docs: apps, compute.

For the parent category (profiles, primitives, governance, BYO runtime), read What is AI agent infrastructure?.

Verified against platform defaults and @usenaive-sdk/server types as of 2026-07-30 (approvals, apps).

In scope on this page / not yet

Bottom line

AI-native cloud is not a synonym for “we bought GPUs,” and it is not Naïve’s one-liner. It is the hosting slice whose primary user can be an agent: provisioned by API, metered per call, governed like every other dangerous capability — beside identity, money, and revoke on the same profile.