← Back to Ascendipiti How it's built · Part 1 · Architecture
How it's built. Alien intelligence. Human control. Your own AI, as powerful as a company.
The AI-native companies we create run on aBOS by Alien Cortex.
aBOS gives each company its own private AI team, built from one tested design. The AI drafts and prepares. People decide. Anything that cannot be undone, like a payment or a send, waits for a named person. The company can leave with its data and setup at any time.
Every setup has the same six core agent classes. None of them is optional.
Super-agent. Runs the day, hands out the work and brings the big calls to a person.Lifecycle and Commercial. Onboarding, proposals and renewals. It drafts. A person sends.Knowledge and Brain (Curator). Keeps the company memory in order and finds things again.Finance: CFO and Bookkeeper. Keeps the books and watches cash on a computer the business owns. It never moves money.Loop-closer. Chases every open item until it is done or handed back to a person.Growth. Finds demand and writes outreach. Nothing goes out until a person says yes.The aBOS loops End-to-end flow: people reach the AI Audit, the Cell Factory builds each company a cell where six agent classes work, and anything irreversible passes a named-person approval gate before reaching the outside world. Four loops sit on this flow. 1 Governance loop, at the factory and the gate: rules compiled in at creation, agents propose, a named person approves the irreversible, trust rises or falls with the override rate. 2 Self-learning loop, in the cell and at the gate: agent drafts, a person approves or corrects, the correction is captured, memory is updated by sensitivity tier. 3 Self-improving loop, in the cell: propose a better way to work, test only rules people chose, adopt or reject with the reason kept, revisit when the reason stops being true. 4 Compounding loop, at the factory and the shared library: a pattern proven in one cell passes a provenance gate with no customer data, joins the shared library, and the factory starts new agents and cells from it. Only patterns go up, never data. HOW THE SYSTEM RUNS, END TO END People owner, team, clients AI Audit the one front door Cell Factory builds every cell 1 4 Company's cell six agent classes at work 2 3 Approval gate a named person says yes 1 2 Outside world sends, payments, filings Numbered badges show where each loop sits on the flow. 1 Governance loop Rules compiled in at creation Agent proposes Named person approves the irreversible Trust rises or falls with override rate Rules are built into each cell when it is created, not set later. Agents propose. A named person approves anything irreversible. An agent earns more freedom only with a low, falling correction rate and passing tests. 2 Self-learning loop Agent drafts Person approves or corrects Correction captured at that moment Memory updated, by sensitivity tier Every correction is captured when it happens. The override rate, the share of work a person had to change, shows whether the system is learning this business. Finance data never enters general memory. 3 Self-improving loop Propose a better way to work Test only rules people chose Adopt, or reject with the reason kept Revisit when a reason stops being true The system improves how the work is done, not only the work. It never tests changes to laws, contracts or safety limits, and it keeps every rejected idea with its reason. 4 Compounding loop Pattern proven in one cell Provenance gate: no customer data Shared library Factory starts new agents and cells from it What one company's agents learn about a class of job goes to the shared library, and every other company benefits. Only patterns cross, never data. If ancestry cannot be proven, it does not cross. override rate sets trust captured corrections proven patterns HOW IT GROWS The factory keeps stamping cells. The shared library keeps compounding patterns. Cell Factory: one core release plus an exact pack lock stamps Cell Cell Cell Cell Cell Next cell only patterns go up, never customer data Shared library of proven patterns new cells start from proven patterns New agents and new cells start from patterns already proven elsewhere, so improvements compound instead of adding up one company at a time. The aBOS loops End-to-end flow: people reach the AI Audit, the Cell Factory builds each company a cell where six agent classes work, and anything irreversible passes a named-person approval gate before reaching the outside world. Four loops sit on this flow. 1 Governance loop, at the factory and the gate: rules compiled in at creation, agents propose, a named person approves the irreversible, trust rises or falls with the override rate. 2 Self-learning loop, in the cell and at the gate: agent drafts, a person approves or corrects, the correction is captured, memory is updated by sensitivity tier. 3 Self-improving loop, in the cell: propose a better way to work, test only rules people chose, adopt or reject with the reason kept, revisit when the reason stops being true. 4 Compounding loop, at the factory and the shared library: a pattern proven in one cell passes a provenance gate with no customer data, joins the shared library, and the factory starts new agents and cells from it. Only patterns go up, never data. HOW THE SYSTEM RUNS, END TO END People owner, team, clients AI Audit the one front door Cell Factory builds every cell 1 4 Company's cell six agent classes at work 2 3 Approval gate a named person says yes 1 2 Outside world sends, payments, filings Numbered badges show where each loop sits on the flow. 1 Governance loop Rules compiled in at creation Agent proposes Named person approves the irreversible Trust rises or falls with override rate Rules are built into each cell when it is created, not set later. Agents propose. A named person approves anything irreversible. An agent earns more freedom only with a low, falling correction rate and passing tests. 2 Self-learning loop Agent drafts Person approves or corrects Correction captured at that moment Memory updated, by sensitivity tier Every correction is captured when it happens. The override rate, the share of work a person had to change, shows whether the system is learning this business. Finance data never enters general memory. override rate sets trust (loop 1), captured corrections feed loop 3 3 Self-improving loop Propose a better way to work Test only rules people chose Adopt, or reject with the reason kept Revisit when a reason stops being true The system improves how the work is done, not only the work. It never tests changes to laws, contracts or safety limits, and it keeps every rejected idea with its reason. proven patterns feed loop 4 4 Compounding loop Pattern proven in one cell Provenance gate: no customer data Shared library Factory starts new agents and cells from it What one company's agents learn about a class of job goes to the shared library, and every other company benefits. Only patterns cross, never data. If ancestry cannot be proven, it does not cross. HOW IT GROWS The factory keeps stamping cells. The shared library keeps compounding patterns. Cell Factory: one core release plus an exact pack lock Cell Cell Cell Next cell only patterns go up, never customer data Shared library of proven patterns New agents and new cells start from patterns already proven elsewhere, so improvements compound instead of adding up one company at a time. Open the full architecture map · Download HTML · Download PDF
How it's built · Part 2 · Updates and journey
Build log Build log entries start here soon. We will post short, dated notes on what changed and why.