Hosted vs Bring-Your-Own Runtime for AI Agents

Where an AI agent runs and how its actions are governed are separate questions. Naïve is runtime-agnostic: inject governed tools into your existing loop (bring your own) or optionally let Naïve operate microVMs (hosted). Either way, the governance gateway enforces spend caps, approvals, audit, and revoke at the tool-call boundary — self-hosting does not bypass policy.

This post compares both paths so you can choose without re-architecting the governed agent profile.

Why runtime-agnostic matters

Teams stall on agent infrastructure because they're asked to migrate the entire harness — swap LangGraph for a vendor runtime, rewrite tool calling, re-test every edge case. Naïve's lead path is the opposite: keep your loop, add real-world capabilities (identity, money, comms) as governed tools.

Governance lives in the gateway, not the runtime. That means Eve on Vercel, AgentCore on AWS, and a Naïve-hosted microVM all hit the same caps and approval gates when they call cards.charge.

Path A — Bring your own runtime (lead path)

Lowest friction if you already ship agents:

import { Naive } from "@usenaive-sdk/server";
const naive = new Naive({ apiKey: process.env.NAIVE_SECRET_KEY! });
 
const op = await naive.forUser(agentId).provision("support", { idempotencyKey: `op:${agentId}` });
const naiveTools = await op.tools();
 
await theirAgent.run({
  tools: [...theirInternalTools, ...naiveTools],
  task: "Resolve ticket #492. Refund if policy allows.",
});

A refund still routes through the gateway — capped, approval-routed above threshold, audited, instantly revocable. See introducing the Naïve SDK and launching the developer platform.

Authenticate the connection safely: mint a short-lived MCP session per tenant instead of embedding your workspace key in the runtime.

When BYO fits

• You already run LangGraph, CrewAI, custom Python, or a vendor agent framework.
• You want governed real-world actions this quarter, not a runtime migration.
• Your customers host agents in their own cloud and you provide the policy layer.

Path B — Naïve-hosted runtime (optional)

Declare a pool, start agents on isolated containers:

await naive.runtime("pool").start(op.id, { goal: "Triage support email." });

Inside a hosted agent, use @usenaive-sdk/runtime:

import { agentProfile } from "@usenaive-sdk/runtime";
const tools = agentProfile.tools();
await runAgentLoop({ tools, goal: "Resolve tickets. Refunds over $50 route to a human." });

Credentials are injected at boot — never on the agent's filesystem. The container is sealed to a per-agent-profile scoped key, not your company key.

Hosted runtime is enabled per deployment. When not configured, start() returns a clear error directing you to BYO — governance is identical either way.

Hosted billing note

A claimed hosted-runtime container is duration-metered: 2 credits per vCPU-hour plus 0.25 credits per GB-hour (per published docs). Default sizing is 1 vCPU / 2 GB. BYO incurs no Naïve runtime metering — you pay your own compute; Naïve meters the governed actions.

When hosted fits

• You want Naïve to operate agent compute without building a Firecracker fleet yourself.
• Multi-agent systems with a declared topology (startSystem) and shared budget.
• You prefer credentials injected by the platform over managing runtime env vars.

Comparison

Two planes, one policy

Naïve separates business-action governance (budgets, approvals, audit, revoke — the differentiated layer) from system governance (microVM network/filesystem isolation on hosted paths). BYO adopters get the business layer immediately; hosted adds the system layer when you opt in.

The governance gateway docs explain how policy maps to Account Kits, approvals, and revoke.

A practical adoption sequence

1. BYO + one primitive — provision a profile, inject op.tools(), gate card issue on approval.
2. Sessions — replace workspace keys in MCP clients with per-user sessions.
3. Hosted (if needed) — declare a pool when you want Naïve to run the loop, not just govern it.

If you're building into multi-tenant SaaS, BYO per customer runtime + Naïve governance is usually the first production shape.

Where to go next

Pick BYO unless you have a concrete reason for hosted compute now. The runtime docs cover pools, startSystem, and BYO examples. The governed agent profile shows how runtime choice fits the full lifecycle.