AI Stack Library · Start here
What an AI operating system for business is, and the twelve layers it is built from
An AI operating system for business is the set of models, agents, memory, rules and schedules that does real work for a company, on a machine the company controls, with a person approving anything that leaves the building.
| If you are | Start with |
|---|---|
| One person running the business from a laptop and a few cloud apps | A dedicated always-on machine for the agent, so your laptop can close; then a rules file, a memory folder and cloud models |
| A small team whose work touches private client data | A dedicated always-on machine with local models for the private jobs, and an approval step before anything is sent |
| A developer building agents for client businesses | A dedicated always-on machine for each client, then the harness and permissions layers, because that is where a client's rules get enforced |
Every business starts with a dedicated always-on machine
An agent that lives on your laptop stops when the lid closes. A dedicated machine keeps the schedules running at three in the morning, keeps your business files in one place you own, and keeps the agent's accounts and keys off the computer you use every day.
Every Hatched Stack runs on one, a machine the business owns.
The twelve layers, in the order you build them
- Keep it on: a machine that stays awake and restarts its services by itself.
- Reach it without a screen: remote access, so the box can sit in a closet.
- Sign-in and disk encryption: decided on purpose, because each choice changes what survives a restart.
- Accounts, logins and keys: held in a vault, never pasted into a chat.
- Models, local and cloud: picked per job, not one model for everything.
- The harness: the program that runs the agent, its tools and its permissions.
- Memory that survives: plain files and an index the agent reads before it works.
- Rules, written once: one file every agent reads at the start of a session.
- Tools and skills: written procedures the agent loads instead of guessing.
- Permissions that are enforced, not asked for: checks that stop a risky command before it runs.
- Schedules and pipelines: the jobs that run on a timer, such as the nightly backup or the morning report.
- Watch it, back it up, keep it from filling up: the layer that tells you something broke.
What to own and what to rent
Own the files that grow more useful every week: your context, your rules, your skills, your data and your logs. They move with you when you change model providers.
Rent the horsepower that someone else runs better at scale: frontier models, email delivery, hosting and offsite backup.
Questions
Is an AI operating system a replacement for Windows or macOS?
No. It runs on top of an ordinary computer. The name describes the job it does for the business, not the software that boots the machine.
Does an AI operating system need local models?
No. You can run every job on cloud models and add local models later for private or high-volume work.
Sources
- Model Context Protocol, checked 2026-10-20
- AGENTS.md, checked 2026-10-20
- Ollama, checked 2026-10-20
- Anthropic: Building effective agents, checked 2026-10-20