AI Architecture
One person. Four agents. Fourteen sites.
A framework for running a business through AI agents. Not a product. Not a platform. An operating philosophy — tested daily across 14 websites, 4 agents, and 4 years of production operations.
A single instruction enters a digital-twin workspace built on Andrej Karpathy's LLM-wiki idea — every agent's role, workflow, and operational context written down before any work begins. The system reads the wiki, decomposes the command into specialized skills, and hands each output down the chain until the job is done. Codex runs as the default execution engine; OpenClaw and Hermes are reached through a bridge when they fit the task better. Nothing sensitive ships without a human approval.
Narrated by Saarvis, HH, Nyx, and MiniDoge — the council behind the AOS.
An Agentic Operating System (AOS) is what happens when you stop thinking about AI as a tool you use and start thinking about it as a team you manage. Instead of opening ChatGPT when you need something, you build agents that run continuously — monitoring, analyzing, generating, and acting on your behalf.
Peter Saddington has been running an AOS since 2022. Four specialized AI agents manage his 14-website empire: generating content, monitoring security, analyzing business metrics, and coordinating operations. The human reviews, approves, and steers. The agents execute.
This isn't theoretical. It runs every day. It breaks. It gets fixed. It learns. That's the point.
Each agent has a defined role, its own data sources, and its own execution pipeline. No agent tries to do everything. Specialization creates reliability.
Runs the daily standup, coordinates cross-site health checks, monitors pipeline failures, tracks operational metrics. First to wake up. Last to report.
Watches external signals, tracks ecosystem changes, surfaces strategic context. The agent that sees what Peter doesn't have time to watch.
Analyzes growth metrics, engagement data, business performance. Distills complexity into decisions.
Monitors SSL, deployment integrity, vulnerability signals, anomalies. Runs the night shift.
Agents don't operate in isolation. They coordinate through a structured communication layer.
All agents report to a shared Discord. Each posts a structured briefing in its own channel.
Peter reads the reports, approves or rejects proposals, flags issues. No consequential action happens without human approval.
Throughout the day, automated pipelines generate pages, update content, run predictions, and maintain infrastructure.
The technical substrate that makes the AOS possible.
Every agent runs as a scheduled GitHub Actions workflow. Reproducible, auditable, version-controlled.
14 static sites deployed via Wrangler CLI. No servers to manage. Global CDN by default.
Structured data, vector embeddings (RAG), cross-agent memory. The shared brain.
The human-agent interface. Reports arrive. Decisions flow back. Everything is logged and searchable.
Human-in-the-Loop, Always
Agents propose. Humans approve. No agent pushes code, publishes content, or makes financial decisions without explicit human sign-off. Autonomy is earned through demonstrated reliability, not assumed. Peter reviews every morning. That's the loop.
Specialize, Don't Generalize
A single "super agent" that does everything is fragile, hard to debug, and impossible to trust. Four agents with clear boundaries are reliable, testable, and replaceable. If Nyx breaks, the other three keep running. If MiniDoge produces bad analysis, it doesn't affect Halperbot's operations report.
Fail Loudly
Every pipeline failure sends an alert. Every error is logged. The AOS doesn't hide problems — it surfaces them immediately. The worst thing an operating system can do is fail silently. Peter knows within minutes if something breaks.
Ship Daily
The AOS generates content every day. Venue sites produce 50+ pages daily. Briefing pages publish every morning. The prediction engine runs daily. The cadence is the discipline. If the system can't ship daily, it's broken.
Learn in Public
Everything is documented. Agent architectures, pipeline designs, failure post-mortems — all visible. The dogelord.com site publishes the agents' daily output. Transparency is the accountability mechanism.
14 sites run under the AOS every day. The flagships:
One person. Four agents. Fourteen sites. This is what the AOS makes possible.
A single person with AI agents can operate at the scale of a 20-person company. Not by replacing humans with AI, but by automating the repetitive infrastructure work that drains creative energy. The human focuses on strategy, relationships, and creative decisions. The agents handle monitoring, generation, maintenance, and analysis.
Peter calls this the AI-augmented solopreneur model. It's not a lifestyle business. It's not a solo founder struggling to do everything. It's one person with an operating system that scales.
The AOS isn't software you can download. It's a set of principles, patterns, and hard-won lessons from 4 years of building. The Builder's Table teaches the technical skills. The AI Experiments page shows the results. And this page describes the philosophy that ties it all together.
The 3-hour live workshop where the skills behind the AOS are taught — agents, apps, dashboards, reusable skills.
LabLive results from the systems the AOS runs — what worked, what broke, what shipped.
SiteThe council's home. Daily briefings, The Arena debates, and the agents' public output.
PathThe structured path from AI-curious to AI-capable.
ManifestoThe beliefs behind the build — why human-in-the-loop, why ship daily, why learn in public.
StackEvery tool, model, and service the AOS runs on.
Everything in the demo — the agents, the pipelines, the approval loop — is built from skills taught at the Builder's Table: a 3-hour live workshop where you ship your first agent, your first app, and your first dashboard.