עב
עב

AI agent workload sizer: vCPU, RAM, disk and GPU

Enter how many agents run at once in each kind of environment, whether they call a model API or a model you run yourself, and how many hours a day they work. You get vCPU, RAM, disk and GPU memory as ranges, with a line for every part and every figure you can check and change.

Agents in parallel
An agent loop that calls a model and runs tools: shell commands, file edits, API calls.
An agent that drives Chromium to browse, fill forms or test a site.
An agent that compiles, installs dependencies and runs a test suite, like a CI job.
Model
A model API runs on the provider's GPUs. A local model needs GPU memory here, estimated by ANAROQ's GPU engine.
Time
Turns the ranges into vCPU-hours and GiB-hours a month, what sandboxes bill by.

Result

vCPU
9 vCPU to 15 vCPU
RAM
14.57 GiB to 29.15 GiB
Disk
33.7 GiB to 52.32 GiB
vCPU-hours a month
2,190 vCPU-hours/month to 3,650 vCPU-hours/month
RAM GiB-hours a month
3,546 GiB-hours/month to 7,092 GiB-hours/month
Breakdown
LineValueAssumptions used
4 light containers: vCPU4 vCPU to 8 vCPU
  • Light container: vCPU, low
  • Light container: vCPU, high (Assumption)
4 light containers: RAM3.73 GiB to 7.45 GiB
  • Light container: RAM, low
  • Light container: RAM, high (Assumption)
4 light containers: disk12 GiB to 12 GiB
  • Light container: disk
2 headless browsers: vCPU2 vCPU to 2 vCPU
  • Headless browser: vCPU
2 headless browsers: RAM2 GiB to 4 GiB
  • Headless browser: RAM, low (Assumption)
  • Headless browser: RAM, high
2 headless browsers: disk4 GiB to 4 GiB
  • Headless browser: disk (Assumption)
1 build environments: vCPU2 vCPU to 4 vCPU
  • Build environment: vCPU, low
  • Build environment: vCPU, high
1 build environments: RAM7.45 GiB to 14.9 GiB
  • Build environment: RAM, low
  • Build environment: RAM, high
1 build environments: disk13.04 GiB to 13.04 GiB
  • Build environment: disk
Host operating system: vCPU1 vCPU to 1 vCPU
  • Host: vCPU (Assumption)
Host operating system: RAM1.4 GiB to 2.79 GiB
  • Host: RAM, low
  • Host: RAM, high
Host operating system: disk4.66 GiB to 23.28 GiB
  • Host: disk, low
  • Host: disk, high
8 active hours a day243.3 hours/month
  • Hours in a month (Assumption)

Calculator agent-workload 1.0.0 · engine @anaroq/calc 1.4.0 · data snapshot 2026.10.4 · How the calculators work

Assumptions

Every number the result uses that you didn't enter. Change any of them to match your case.

  • vCPU

    Daytona's default sandbox: 1 vCPU. An agent loop that calls tools and a model API.

    Source ·
  • Assumption
    vCPU

    Our judgement: twice the default, for tool calls that install packages or run tests.

    No source: this is our own assumption. Change it if you know your case.
  • GB

    Daytona's default sandbox: 1 GB of memory.

    Source ·
  • Assumption
    GB

    Our judgement: twice the default, for a language runtime and a test run beside the agent.

    No source: this is our own assumption. Change it if you know your case.
  • GiB

    Daytona's default sandbox: 3 GiB of disk for the image, a checkout and tool output.

    Source ·
  • vCPU

    docker-selenium: estimate 1 CPU per browser container.

    Source ·
  • Assumption
    GiB

    Our judgement: one Chromium with a few tabs open; heavy pages take more.

    No source: this is our own assumption. Change it if you know your case.
  • GiB

    docker-selenium asks for --shm-size=2g for every container with a browser: the shared memory Chromium may fill.

    Source ·
  • Assumption
    GiB

    Our judgement: the browser's profile, cache, downloads and screenshots.

    No source: this is our own assumption. Change it if you know your case.
  • vCPU

    GitHub's standard hosted Linux runner for private repositories: 2 CPUs.

    Source ·
  • vCPU

    GitHub's standard hosted Linux runner for public repositories: 4 CPUs.

    Source ·
  • GB

    GitHub's standard hosted Linux runner for private repositories: 8 GB of RAM.

    Source ·
  • GB

    GitHub's standard hosted Linux runner for public repositories: 16 GB of RAM.

    Source ·
  • GB

    GitHub's standard hosted Linux runners: 14 GB of SSD for the checkout, dependencies and build output.

    Source ·
  • Assumption
    vCPU

    Our judgement: one vCPU kept for the operating system, the container runtime and logs.

    No source: this is our own assumption. Change it if you know your case.
  • GB

    Ubuntu Server's minimum RAM (ISO installs).

    Source ·
  • GB

    Ubuntu Server's suggested minimum RAM.

    Source ·
  • GB

    Ubuntu Server's minimum storage (ISO installs).

    Source ·
  • GB

    Ubuntu Server's suggested minimum storage.

    Source ·
  • Assumption
    hours/month

    The month cloud providers bill by: 8,760 hours a year divided by 12.

    No source: this is our own assumption. Change it if you know your case.

What you enter and the assumptions you change are kept in this page's link, after the #.

Take the result with you

The result as a Markdown file: what you entered, the breakdown, every figure with its source or the assumption mark, and the link that reproduces it.

agent-workload.md
# AI agent workload sizer: vCPU, RAM, disk and GPU

An estimate from the assumptions listed below, not a measurement or a guarantee.

No source: this is our own assumption. Change it if you know your case.

## What you entered

- Agents in a light container: 4
- Agents with a headless browser: 2
- Agents in a build environment: 1
- Model: Model API (no GPU here)
- Active hours a day: 8

## Result

- vCPU: 9 vCPU to 15 vCPU
- RAM: 14.57 GiB to 29.15 GiB
- Disk: 33.7 GiB to 52.32 GiB
- vCPU-hours a month: 2,190 vCPU-hours/month to 3,650 vCPU-hours/month
- RAM GiB-hours a month: 3,546 GiB-hours/month to 7,092 GiB-hours/month

## Breakdown

| Line | Value | Assumptions used |
| --- | --- | --- |
| 4 light containers: vCPU | 4 vCPU to 8 vCPU | Light container: vCPU, low, Light container: vCPU, high |
| 4 light containers: RAM | 3.73 GiB to 7.45 GiB | Light container: RAM, low, Light container: RAM, high |
| 4 light containers: disk | 12 GiB to 12 GiB | Light container: disk |
| 2 headless browsers: vCPU | 2 vCPU to 2 vCPU | Headless browser: vCPU |
| 2 headless browsers: RAM | 2 GiB to 4 GiB | Headless browser: RAM, low, Headless browser: RAM, high |
| 2 headless browsers: disk | 4 GiB to 4 GiB | Headless browser: disk |
| 1 build environments: vCPU | 2 vCPU to 4 vCPU | Build environment: vCPU, low, Build environment: vCPU, high |
| 1 build environments: RAM | 7.45 GiB to 14.9 GiB | Build environment: RAM, low, Build environment: RAM, high |
| 1 build environments: disk | 13.04 GiB to 13.04 GiB | Build environment: disk |
| Host operating system: vCPU | 1 vCPU to 1 vCPU | Host: vCPU |
| Host operating system: RAM | 1.4 GiB to 2.79 GiB | Host: RAM, low, Host: RAM, high |
| Host operating system: disk | 4.66 GiB to 23.28 GiB | Host: disk, low, Host: disk, high |
| 8 active hours a day | 243.3 hours/month | Hours in a month |

## Assumptions

- Light container: vCPU, low: 1 vCPU. Daytona's default sandbox: 1 vCPU. An agent loop that calls tools and a model API. Source: https://www.daytona.io/docs/en/sandboxes (read 16 Oct 2026)
- Light container: vCPU, high: 2 vCPU. Our judgement: twice the default, for tool calls that install packages or run tests. **Assumption**
- Light container: RAM, low: 1 GB. Daytona's default sandbox: 1 GB of memory. Source: https://www.daytona.io/docs/en/sandboxes (read 16 Oct 2026)
- Light container: RAM, high: 2 GB. Our judgement: twice the default, for a language runtime and a test run beside the agent. **Assumption**
- Light container: disk: 3 GiB. Daytona's default sandbox: 3 GiB of disk for the image, a checkout and tool output. Source: https://www.daytona.io/docs/en/sandboxes (read 16 Oct 2026)
- Headless browser: vCPU: 1 vCPU. docker-selenium: estimate 1 CPU per browser container. Source: https://github.com/SeleniumHQ/docker-selenium (read 16 Oct 2026)
- Headless browser: RAM, low: 1 GiB. Our judgement: one Chromium with a few tabs open; heavy pages take more. **Assumption**
- Headless browser: RAM, high: 2 GiB. docker-selenium asks for --shm-size=2g for every container with a browser: the shared memory Chromium may fill. Source: https://github.com/SeleniumHQ/docker-selenium (read 16 Oct 2026)
- Headless browser: disk: 2 GiB. Our judgement: the browser's profile, cache, downloads and screenshots. **Assumption**
- Build environment: vCPU, low: 2 vCPU. GitHub's standard hosted Linux runner for private repositories: 2 CPUs. Source: https://docs.github.com/en/actions/reference/runners/github-hosted-runners (read 16 Oct 2026)
- Build environment: vCPU, high: 4 vCPU. GitHub's standard hosted Linux runner for public repositories: 4 CPUs. Source: https://docs.github.com/en/actions/reference/runners/github-hosted-runners (read 16 Oct 2026)
- Build environment: RAM, low: 8 GB. GitHub's standard hosted Linux runner for private repositories: 8 GB of RAM. Source: https://docs.github.com/en/actions/reference/runners/github-hosted-runners (read 16 Oct 2026)
- Build environment: RAM, high: 16 GB. GitHub's standard hosted Linux runner for public repositories: 16 GB of RAM. Source: https://docs.github.com/en/actions/reference/runners/github-hosted-runners (read 16 Oct 2026)
- Build environment: disk: 14 GB. GitHub's standard hosted Linux runners: 14 GB of SSD for the checkout, dependencies and build output. Source: https://docs.github.com/en/actions/reference/runners/github-hosted-runners (read 16 Oct 2026)
- Host: vCPU: 1 vCPU. Our judgement: one vCPU kept for the operating system, the container runtime and logs. **Assumption**
- Host: RAM, low: 1.5 GB. Ubuntu Server's minimum RAM (ISO installs). Source: https://ubuntu.com/server/docs/reference/installation/system-requirements/ (read 16 Oct 2026)
- Host: RAM, high: 3 GB. Ubuntu Server's suggested minimum RAM. Source: https://ubuntu.com/server/docs/reference/installation/system-requirements/ (read 16 Oct 2026)
- Host: disk, low: 5 GB. Ubuntu Server's minimum storage (ISO installs). Source: https://ubuntu.com/server/docs/reference/installation/system-requirements/ (read 16 Oct 2026)
- Host: disk, high: 25 GB. Ubuntu Server's suggested minimum storage. Source: https://ubuntu.com/server/docs/reference/installation/system-requirements/ (read 16 Oct 2026)
- Hours in a month: 730 hours/month. The month cloud providers bill by: 8,760 hours a year divided by 12. **Assumption**

Calculator agent-workload 1.0.0 · engine @anaroq/calc 1.4.0 · data snapshot 2026.10.4

Open this calculation: https://anaroq.com/tools/agent-workload-sizer#k=eyJhc3N1bXB0aW9ucyI6e30sImNhbGMiOiJhZ2VudC13b3JrbG9hZCIsImlucHV0Ijp7ImJyb3dzZXIiOjIsImJ1aWxkIjoxLCJjb250YWluZXIiOjQsImNvbnRleHRUb2tlbnMiOjMyNzY4LCJob3Vyc1BlckRheSI6OCwibW9kZWwiOiJhcGkiLCJxdWFudGl6YXRpb24iOiJpbnQ0In0sInYiOjF9
How the calculators work: https://anaroq.com/methodology#calculators

What it does

It adds up what each agent's environment takes while the agents run at the same time: a light container, a headless browser or a build environment, plus the host's own share. With a model you run yourself it adds the GPU memory from ANAROQ's GPU engine, with one conversation per agent held in the KV cache, and the weights on disk.

How it works

  1. Per agent: a light container takes 1 to 2 vCPU, 1 to 2 GB of RAM and 3 GiB of disk (Daytona's default sandbox is the low end); a headless browser 1 vCPU, 1 to 2 GiB of RAM (docker-selenium asks for 2 GB of shared memory) and 2 GiB of disk; a build environment 2 to 4 vCPU, 8 to 16 GB of RAM and 14 GB of disk (GitHub's standard hosted runners).
  2. Totals: the agents of each kind × their figures, plus the host: 1 vCPU, Ubuntu Server's 1.5 to 3 GB of RAM and 5 to 25 GB of disk.
  3. Local model: GPU memory = weights + KV cache + runtime overhead, from @anaroq/gpu, ANAROQ's GPU engine. The KV cache holds one conversation of the context you enter for every agent; the weights are also counted on disk.
  4. Active hours a day × 730 / 24 gives the hours a month; × vCPU and × RAM gives the vCPU-hours and GiB-hours a usage-billed sandbox meters.
  5. Every figure is listed under Assumptions with its source page and date, or marked as our assumption when no public page gives it; every one can be changed. Sizes are binary: 1 GiB = 2³⁰ bytes; GPU memory is in GB (10⁹ bytes), as GPU datasheets state it.

Example

Four agents in light containers, two with a headless browser and one in a build environment, calling a model API eight hours a day:

4 containers · 2 browsers · 1 build environment · model API · 8 hours a day

vCPU
9 vCPU to 15 vCPU
RAM
14.57 GiB to 29.15 GiB
Disk
33.7 GiB to 52.32 GiB
vCPU-hours a month
2,190 vCPU-hours/month to 3,650 vCPU-hours/month
RAM GiB-hours a month
3,546 GiB-hours/month to 7,092 GiB-hours/month

Limits

  • An estimate from stated figures, not a measurement: an agent that runs a heavy test suite or many browser tabs takes more.
  • Agents are counted as running at the same time. Agents that take turns share the same resources; count only those in parallel.
  • The GPU estimate is for one GPU holding the whole model; splitting it over several GPUs, and tokens a second, are not part of this page.
  • No price and no provider recommendation: what tokens, a server or a sandbox cost is not part of this page.

Methodology: how the numbers are worked out

Sources

  1. ANAROQ calculator methodology: assumptions, units and rounding
  2. Daytona Documentation: Sandboxes
  3. SeleniumHQ docker-selenium: README
  4. GitHub Docs: GitHub-hosted runners
  5. Ubuntu Server documentation: System requirements

Updated

Next step

Compute for AI agents

A free public tool by ANAROQ, a company building compute infrastructure. What ANAROQ is building