How I Became an AI Orchestrator: 90 Days from Zero to Agency

Background#

Engineer since 15. Then 10 years in e-commerce – DTC brands, teams of 20 to 60, scaling from $1M to eight figures.

When AI agents appeared – I dove in. Not because “the market is shaking” or “time for a career change.” Because it’s genuinely fun, like a game. A new instrument that lets you build things that were impossible before. And people pay for it.

Month 1: Terminal, Chaos, First Skills#

The first two weeks were honest chaos. ChatGPT, Claude, dozens of API keys, prompts in a notepad. Three in the morning, five terminals open, nothing works together. Plenty of power – no structure. A Ferrari without a steering wheel.

Then I started working with Claude Code – from day one of its launch. At first like everyone else: tasks, code, features. Typical vibe coding. But gradually I saw something deeper: this isn’t just an assistant. Through AI you can build entire programs, entire releases – by voice. Without writing code by hand.

Then OpenClaw appeared. I joined from day one.

OpenClaw is an open-source platform that turns an AI model into an autonomous agent. Not SaaS. Not a subscription. Local server, your data, your control. I set it up in one evening and started building on top of it.

The first thing that worked – morning briefings. Every morning the agent compiles a summary: what happened overnight, what’s on the schedule, what needs attention. I used to spend 40 minutes on this. Now – zero. I just read the finished brief in Telegram.

Second – memory. I connected SurrealDB and started building a system where the agent remembers context permanently – not per session, but always. Nightly consolidation – the agent decides what’s worth remembering. Vector search across extracted facts. Over time I added entities – digital profiles of people, concepts, and projects that you can query directly.

By the end of month one it was clear: an AI agent isn’t a tool. It’s an employee you need to train. And I have 10 years of experience training teams.

Month 2: Products and Real Problems#

Every day I added skills – modular capabilities for the agent. Downloaded over 150 from the marketplace. Wrote about 20 myself. Twitter monitoring, Telegram publishing, health data analysis with Oura, calendar sync.

X-Intel – I built this Twitter/X intelligence system myself. The agent monitors accounts that matter to me, scores signals by relevance, and delivers a digest every morning. What used to take 2-3 hours of scrolling – now 15 minutes of reading in Telegram.

Ember – voice bridge. Two buttons on a Logitech MX Master 3 mouse: press – speak – release – the agent has the task. No keyboard, no window switching.

Gnosis Panel – admin dashboard for managing the whole system. Tasks, skills, system health – everything in one place.

But the biggest lesson of month two came from problems.

The gateway crashed. Restart loops – bloated sessions plus wrong timeouts. The system spent more time restarting than working. I built multi-layered protection: model fallback chains, watchdog, circuit breaker, disk monitoring, automatic session cleanup. Each layer – a response to a specific incident I lived through.

Security needed hands-on fixing too. Keys ended up in the wrong places. I set up proper storage, wrapper scripts, validation before every launch.

I’ve been an engineer-programmer since I was 15. Sites, programs, systems – that’s my background long before e-commerce. Code now gets written by my AI – but I design the architecture. I define agent roles. I diagnose problems. When the system goes down at night – I’m the one figuring out why. AI speeds up the hands. The head is mine.

Month 3: From My System to Systems for Others#

By the third month I realized: what I built for myself – others need too.

I wrote an installer – a script that deploys the full system on a VPS in 7-10 minutes. Configuration templates, smoke tests, deployment checklist. Each client gets a ready system, not one assembled from scratch.

First clients came from my circle. Entrepreneurs – people who genuinely benefit from optimizing their time. And just friends who joined the game and now develop their own bots.

Over 90 days – 7 clients. Los Angeles and remote. Each one gets personalized setup: I study their processes, build the system, train the agent for their context.

I set up a separate server for my parents with two agents – one for dad, one for mom. In parallel I packaged my configuration into Gnosis OS – so I could deploy to new clients in hours, not weeks.

How I Work Now#

Two instruments in parallel:

Claude Code – for engineering work. Architecture, code, configuration, debugging. When something needs building or fixing – I’m here.

OpenClaw – for operational work. Research, knowledge base, monitoring, client systems. Agents run 24/7, execute cron jobs, respond in Telegram.

I orchestrate both. That’s what an AI conductor is – not a programmer, not a prompt engineer, but someone who orchestrates AI systems. I tested Hermes, OpenClaw, Claude Code – and settled on a hybrid.

Now I’m building an AI agency. Automation for companies – local in LA and remote. Not a mass product. Personalized setup for each client.

What I Tested in 90 Days#

• Models: Anthropic (Opus, Sonnet, Haiku), Google (Gemini), OpenAI (GPT-5.4), Qwen 3.6 via OpenRouter (free), Groq inference. Dozens of models, dozens of providers. Each one for its own task.
• Harnesses: Hermes, OpenClaw, Claude Code. Dug into how agent loops, tool calling, and context management work under the hood. Currently tracking the fastest-growing repositories in the space.
• Platforms: OpenClaw + Claude Code hybrid. No-code automation isn’t for me – I need control at the code level.
• Databases: SurrealDB – vector search, full-text indexes, graph relationships. One database for knowledge, memory, and analytics.
• Voice: ElevenLabs, Groq STT, Ember – my voice bridge for controlling agents straight from the mouse.

Every day I read what’s new in AI. Watch the market. Test. Break. Fix. This isn’t a hobby – it’s a profession.

Cognitive Sovereignty#

That’s what I call it. Not “AI automates tasks.” But “AI frees the brain for what actually matters.” Strategy, relationships, decisions – that stays with you. Routine, monitoring, data collection, initial analysis – that’s the agent’s job.

Who This Is For#

You use ChatGPT and feel the ceiling – it’s real. But the ceiling isn’t AI’s. It’s the tool you’re using.

For founders drowning in operations: an AI agent handles 60% of the routine. Right now, not next year.

For those curious about running their own AI on their own server – OpenClaw is free, installation takes 15 minutes.

And for those who want to skip my 90 days of trial and error – that’s why I’m building an AI agency.

Want to try it yourself? How to Install OpenClaw: AI Agent in 15 Minutes.

Want to see the system in action? Gnosis OS – Personal AI System 24/7.

Want your own Jarvis? See pricing.

Paid consultations – $200/hour. We’ll get on a call, I’ll show you and teach you. Book a call.

I’ll keep sharing and demonstrating what these new technologies can do. Follow along. Want a system like this – let’s talk. Want training – that works too. I’ll try to publish everything here so anyone can walk this path themselves.