Persistent AI Executives
Agents with permanent identity, role, memory and career history that survive model changes and restarts — the agent runtime is the employee, the LLM is the reasoning engine.
An Operating System for the AI-Operated Company
This project explores a practical research question: can a small, disciplined company be operated by a lean workforce of persistent AI employees — under clear human authority, full auditability, and explicit approval gates?
Keywords: persistent agents · agent identity · multi-agent organisation · governance · auditability · multi-tenancy · multi-jurisdiction compliance · local / air-gapped inference
Summary of the Concept
AI-COS is a web-based operating system in which the CEO remains the ultimate human authority while persistent AI agents perform the majority of departmental work. An agent is not a chatbot with a system prompt: it is a durable digital entity with identity, role, manager, goals, memory, permissions and a career record that survive model changes, restarts and new conversations. The language model is invoked only when the entity needs to reason, decide, communicate or act.
The system begins with a deliberately lean organisation — a Chief of Staff agent and a CTO agent — and grows new capability only when recurring workload and evidence justify it. Everything material is traceable, and nothing autonomous is invisible.
The Agent Is Not the LLM
The persistent Agent Runtime is the employee; the LLM is the reasoning engine the employee uses.
This separation is fundamental to continuity. Identity, history, relationships, responsibilities, skills, experience, permissions and accountability must persist independently of any particular model. An agent can therefore be re-platformed onto a different model — or moved to a local, air-gapped deployment — without losing who it is or what it has learned.
Startup-First, Progressive Organisation
CEO
human authority
Chief of Staff
company coordination
CTO
technology & product
Jobet SaaS
white-label travel OS
GHTE SaaS
halal certification
Jurisdiction 0
Malaysia · SSM/LHDN/PDPA
Jurisdiction N
per-country · subsidiaries
New departments — finance, risk & compliance, marketing — are created only when real workload justifies them, each through a controlled request-and-approval workflow. The platform is enterprise-capable but startup-operated.
Multi-Jurisdiction & Cross-Jurisdiction Operation
AI-COS is built as a multi-jurisdiction, cross-jurisdiction compliance platform. Jurisdiction is a first-class axis of the operating system — not a fixed configuration.
- —Jurisdiction 0 — Malaysia is the starting baseline (SSM, LHDN/HASiL, SST, PDPA) for the platform's own operation.
- —Deployable to multiple countries: each deployment configures its own corporate, tax, e-invoicing and data-protection rule sets from official sources.
- —Subsidiary groups: a company with subsidiaries in other countries operates every subsidiary under its home jurisdiction's obligations from one console.
- —Products are jurisdiction-independent: the same product (e.g. a travel platform or a certification platform) can be deployed per country.
- —Regional / on-premise option: where data-residency requires it, a deployment can run fully local — including the model layer (air-gapped inference).
What the System Enables
No Invisible Autonomy
Every material action is traceable from company level down to an individual tool call and its evidence — the CEO can always answer 'what happened, why, and who is accountable'.
Controlled Autonomous Execution
Autonomy is configurable per agent and per action, from advisory-only to bounded autonomous work, with human approval gates for high-risk or irreversible actions.
Multi-Tenant Isolation
Each company that runs on AI-COS occupies a strictly isolated tenant — data, agents, memory and permissions are separated by design, with operator-level cross-tenant access as the sole exception.
Model-Agnostic Runtime
The same agent can reason through a cloud LLM or a local, air-gapped model — suited to organisations with data-residency or on-premise requirements.
Auditable by Construction
Append-only audit logs, hashed evidence chains, decision records and execution traces are part of the operating system, not an afterthought.
Multi-Jurisdiction by Design
Jurisdiction is a first-class entity: a company — or a group with subsidiaries abroad — is operated under the compliance rules of every country it operates in, with jurisdiction 0 (Malaysia) as the starting baseline.
Deployments Under Study
A multi-tenant travel platform (white-label SaaS) operated by an AI workforce: the Chief of Staff coordinates company priorities and CEO briefings, the CTO owns technical execution across the product roadmap.
- Operational access to the product through its API layer; changes gated by explicit approval.
- Per-tenant AI metering and budgets already provide the cost telemetry the operating layer consumes.
A regulatory-domain SaaS (halal certification, JAKIM-aligned, eight industry sectors) whose audit trail, role-based access and AI-assisted workflows form the evidence base for an advisory AI layer.
- Integration is read-only and owner-authorized; no source sharing across the legal boundary.
- The platform's existing 'AI assists, human decides' principle aligns with the AI-COS governance model.
Logical Layers
| Presentation | CEO cockpit: dashboard, attention queue, approvals, traces | 1 |
| Control Plane | Organization, permissions, policies, approvals, configuration | 2 |
| Agent OS | Identity, context, memory, goals, roles, lifecycle, runtime | 3 |
| Workflow / Events | Tasks, schedules, wakeups, retries, durable execution | 4 |
| Data | Company, agents, memory, work, audit, evidence | 5 |
| Jurisdiction | Per-jurisdiction rules, obligations, calendars, regulator sources | 6 |
| Integrations | GitHub, email, product adapters, external systems | 7 |
| Model | Interchangeable LLM providers — cloud or local / air-gapped | 8 |
Deployment targets Vercel + Supabase/serverless, with an optional on-premise variant in which the model layer runs fully local — suitable for air-gapped or data-residency-constrained organisations. The jurisdiction layer scopes compliance rules, obligations and regulator sources per country, so the same platform serves single-country tenants and multinational subsidiary groups.
Notes & Source Material
- Governance and decision records (multi-tenant isolation, access-vs-change, model policy) are maintained as living documents in the project repository.
- Case-study findings are based on real, running product codebases; findings are presented in sanitized, aggregate form.
- The operating principles follow the “no invisible autonomy” doctrine: every material action is traceable to an actor, a policy, a tool call and evidence.