The Agentic OS Family: One Architecture, Installed Into Real Businesses

TL;DR: I installed a complete AI operating system into a friend's consumer-goods startup in one day: website, agent harness, a zero-tolerance claims gate, and a business memory that compounds weekly. Then I extracted the generic machinery into agentic-business-os, a public MIT template with an update channel that ships harness improvements to every instance as readable pull requests. The install service around it is the FrankX Foundry — application-only, evaluated, with priority for sustainable, healthcare, and meaningful businesses.

What happened this week

Friends of mine are launching a physical product in a category where one careless sentence of marketing copy is a regulatory problem. They asked what my stack could do for them.

The answer turned out to be: everything, in a day. Not because the day was heroic, but because the architecture already existed. frankx.ai runs on a layered system: a memory substrate (Starlight Intelligence System), an agent harness (Agentic Creator OS), a two-file design contract, and pre-publish quality gates. Installing it for a business meant deriving those layers for one brand instead of inventing anything new.

What an installed OS actually is

Five contract files in a repository, read automatically by any coding agent:

• The doctrine (CLAUDE.md): the company handbook, enforced by software
• The design contract (design.md + taste.md): every visual value the site may use, plus the judgment tokens can't capture
• The voice file: banned phrases, blocked claim patterns, approved framing
• The gate (@claims-guard): a pre-publish audit that blocks regulated claim language outright. A FAIL is final until a human rewrites
• The memory (docs/intelligence/): decision records, market notes, weekly reviews

The operating cost once it runs: a ten-minute Monday plan and a fifteen-minute Friday review. The agents do production; the gates hold the floor; the memory makes week 30 smarter than week 1.

The part that makes it a system, not a setup

A setup decays the day it ships. This doesn't, because of one design decision: brand is yours, machinery is shared.

The brand files above are instance-owned — never touched by anyone. The generic machinery (the agents, the command pipelines, the gate logic) is upstream-managed. When it improves in the public template, every registered instance receives a pull request with a plain-language changelog. The founder reads the diff in the GitHub UI and merges or declines. Nothing auto-merges, ever. The contract is one page: HARNESS.md.

The loop is already proven, not promised: the first instance merged its first harness-sync pull request the same week it was installed.

Take the template, or apply for an install

The template and four standalone skill packs (the claims gate, the business memory, the design-contract authoring session, the weekly rhythm) are MIT-licensed at frankxai/agentic-business-os. They work in Claude Code, Cursor, and Codex, and the packs import into a Claude.ai project as plain folders.

If you want it installed, derived for your brand, and connected for the long run: the Foundry takes a small number of installs per quarter, by application, with priority for sustainable, healthcare, and genuinely meaningful products. Pricing follows evaluation; the founding cohort is forming now.

The repos are public on purpose. Verify rather than trust.