What this guide adds
The /cloud/agents page lists the limits to require, and the Agent Sandbox Matrix compares isolation options. This guide adds the measuring: how much compute your agents use at once, what their model calls cost, which risks your setup has, and what machines of that size are listed at in the VPS Price Index.
Before you compare
An AI agent needs two kinds of infrastructure: a place where its code runs in isolation, and the model it calls. Both are sized from the same facts: how many agents run at once, for how long, and what they are allowed to reach.
Every figure here comes from the site's data with its source and read date. The prose names no price and ranks no provider.
What to compare
- Isolation
- Code an agent writes or downloads should run where it cannot reach your other machines or secrets: a container with tight limits, a microVM or a managed sandbox. The stronger the isolation, the less you depend on the agent behaving.
- Concurrency
- Agents are sized by how many run at the same time, not by how many exist. Measure the peak, and cap it on the platform.
- Lifetime
- Short runs suit environments that start fast and expire on their own; long-running agents need a persistent machine, kept with the same care as any server.
- Where the model runs
- An agent that calls a model API needs little compute of its own; one that calls a model you run needs GPU memory. That choice moves most of the cost.
- Limits the platform enforces
- A hard budget, an expiry on every resource, a cap on concurrency, approval for risky actions, an audit trail and a kill switch. A limit written only in a prompt is a request.
- Network reach
- Decide what the environment may reach, such as the model API and a package registry, and deny everything else by default.
Measure with the tools
In this sequence. Each step is answered by a free tool on this site that works in your browser.
- Size the agents
Enter how many agents run at once in each kind of environment, whether they call an API or your own model, and how many hours they work. You get vCPU, memory, disk and GPU memory as ranges.
- Price the model calls
Agents spend tokens. Enter the volume and compare the API you would pay with running an open model on a rented card.
- Assess the risks
Answer questions about where the agent runs and what it can reach. You get findings by severity, each with a recommendation and its source.
- Size a persistent host
For agents that run on a machine you keep, turn the sizer's result and your other services into one server size with headroom.
Tool: Server size calculator
Dated prices from the Price Index
Smaller plans from the VPS Price Index sample, as listed in the dataset, with the vCPUs and memory of each, for comparing with the Agent Workload Sizer's result. Managed sandboxes are not in the index; the Agent Sandbox Matrix compares them.
| Provider | Plan | Region | vCPU | RAM (GiB) | Listed price per month |
|---|---|---|---|---|---|
| DigitalOcean | Basic Droplet (Regular) 1 GiB | One price everywhere | 1 | 1 | $6 |
| DigitalOcean | Basic Droplet (Regular) 4 GiB | One price everywhere | 2 | 4 | $24 |
| Akamai Cloud | Nanode 1 GB (Shared CPU) | North America | 1 | 1 | $5 |
| Akamai Cloud | Linode 4 GB (Shared CPU) | North America | 2 | 4 | $24 |
| AWS | Lightsail 1 GB (Linux/Unix, IPv4) | One price everywhere | 2 | 1 | $7 |
| AWS | Lightsail 4 GB (Linux/Unix, IPv4) | One price everywhere | 2 | 4 | $24 |
| UpCloud | Starter 1 GB | One price everywhere | 1 | 1 | €3 |
| UpCloud | Starter 4 GB, 2 cores | One price everywhere | 2 | 4 | €12 |
Prices as each provider lists them, in its currency, excluding taxes unless the row's source says otherwise. The full index converts them with dated rates.
The full index: VPS price index
Common mistakes
- Agents on your main server
- Running agent code on the machine that holds your data means one mistake reaches everything. Give agents their own environment.
- Sizing for the average
- Agents start in bursts. Size for the peak of concurrent runs, then cap it.
- No expiry
- A forgotten environment keeps running and keeps billing. Give every resource an expiry unless someone keeps it.
- An open network by default
- An environment that can reach anything can leak anything. Allow what the task needs and deny the rest.
- Counting machines, forgetting tokens
- For agents that call an API, the model's tokens can cost more than the machines. Price both.
Limits
- The sizer's figures are estimates from the inputs you give; measure a real run before settling on a size.
- The VPS sample covers a few providers and sizes, and managed sandboxes are not in the Price Index.
- Prices change: every row shows the day it was read and is marked stale after the refresh window. The provider's page is authoritative.
On the /cloud page
The /cloud page for this kind of compute lists what to check before you commit, and describes the principles ANAROQ is building to, with its In development status.