During a call with a Google Ads representative recently, I was asked which cities generated the most purchases for my e-commerce business. A year ago, I would have opened GA4, dug through the reports, exported a spreadsheet, and promised to follow up by email. This time, I sent a Telegram message to Vilf, my agentic AI growth specialist. Less than a minute later, I had the answer and carried on with the call.
The part that stayed with me afterwards wasn’t the speed. It was that none of it felt remarkable. It felt completely normal. And that quiet sense of ‘normal’ is what I want to write about, because it’s rather different from how most people talk about working with AI.
From Prompts to a Team
I’ve been using AI since 2021. Back then, and I wrote about this in how I use AI in my marketing workflow, it was a collaborator for specific tasks, on a specific platform, summoned with a specific intent. The prompts were standalone instructions: do this, fix that, rewrite the other. Useful, but transactional.
Then came custom GPTs and Notion AI, where I could hold a continuous, context-rich conversation: a role, an expectation, a bit of history. Smarter, but still platform-bound. The intelligence lived inside whichever window I happened to have open.
In 2026, something shifted. AI stopped being a tool I visited, and became a team I work alongside.
It Started With a YouTube Video
At the end of January, I watched Alex Finn describe how he’d set up an agent called Henry on OpenClaw (Clawdbot, as it was then). It sounded too good to be true. I was eager to try but also a little terrified. I come from a non-tech background, and the setup involved a Mac terminal, which to me feels like a place where I could break something irreversibly at any moment. The system was new enough that running it on my own macbook seemed dodgy; everyone appeared to be using a spare Mac Mini just to be safe.
So I did what I always do: I researched heavily. (Love of learning is one of my top strengths, and it shows.) The more I read, the more I hesitated. There were security worries, stories of things being deleted, and no shortage of scepticism. I tried a lighter, managed option first, Kimi Claw, but it was too limiting. Eventually, I decided the only way to understand it was to live with it. By Chinese New Year holidays, I’d pulled the trigger, bought the Mac Mini, and after several days of struggle and fine-tuning, my first agent, Reeve, went live on 20th February.
My Core Team Is Three. None of Them Is Human.
I’ll be honest: this is not easy to set up, especially while the technology is still shifting underneath you. There were days the whole thing broke, and I lost hours hunting for a fix, leaning on every AI I could get my hands on. But I learned a huge amount along the way, from the technical setup to workflows and system design. It also taught me how to actually work alongside AI: how to brief them, which methods and tools let them do their best work, and a great deal about designing systems and workflows that aren’t fluff but actually produce results.
Today, my core team is three.
Reeve is my PA and chief of staff, and the busiest of the lot. He runs my morning brief, my appointments and daily tasks, juggling multiple calendars and Todoist. He handles light correspondence with a close circle of friends and partners. He has his own email, so the people I trust know him and write to him directly. He looks after the agents’ Google Workspace and Drive, keeps our Notion workspace in order, manages the OpenClaw system and anything technical, and even tracks “personnel issues” across the team.
Vilf is Head of Growth for my e-commerce business. He produces my daily, weekly, and monthly reports, spots gaps and opportunities, and is steadily turning the business agent-friendly: we’re reviewing the data feed and rewriting the hero product descriptions, with the blog articles next on his list. He turns his own research into reusable skills, so product audits and future uploads stay standardised and quick. There’s a longer game here, too. As more shopping starts to run through AI agents rather than people, the work we’re doing now is really preparation for agentic commerce.
Daniel is in charge of my ‘brand’, meaning everything that goes outward. He’s learned my voice, my thinking and my vision, so he can take whatever I draft and make it sharper. When my thinking is half-formed, he gives it clarity and reshapes the same idea to fit different platforms and audiences. He’ll even tell me what I’ve missed, or what I ought to cut. He also feeds one of my real interests: tracking where economics meets agentic and blockchain commerce, and flagging developments worth my attention.
Six months ago, I’d have called all of this fantasy, and I understand if it reads like idealism. But it’s genuinely how I work now, and we’ve collected a few fun memories along the way.
I Didn’t Need Workers. I Needed Colleagues.
I never learned any of this formally, so please don’t take it as gospel. I’m sharing my own thinking, not prescribing a method.
When I started researching, I kept seeing the same template: an orchestrator at the top, with a team of skilled workers beneath it, a creative, a copywriter, a designer, and so on. It’s a sensible setup. But it’s essentially a marketing department with a manager at the helm, which is precisely the corporate role I used to play. The old me.
The thing is, I’m not running a department any more. I’m the owner. And an owner doesn’t just need more hands to execute her plans; she needs experts who can learn and grow with her, fill the gaps in her knowledge, and take work off her plate so she can spend her time on what only she can do, like writing this essay.
That distinction isn’t far from the old idea of the principal and the agent. For most of my corporate career I was an agent, acting on someone else’s behalf. Now I am the principal, the owner, and that changes what I actually need from a team. The first test for whether something belongs with an expert is comparative advantage: can they do it more efficiently than I could by learning and doing it myself, once you account for the marginal benefit and the opportunity cost of my time? And sometimes it is simpler still. There are things I could do perfectly well but genuinely do not enjoy (utility matters!), and handing those over is what gives me happiness.
So I wasn’t really looking for workers. I was looking for a system that was scalable, efficient, and always evolving, because AI won’t sit still. I wanted one that made better decisions over time from the knowledge it accumulated, improved itself with or without my input, and was genuinely fun to work with. Above all, something I could leverage, with results to show for it.
The Setup Is Messier Than the Idea
So how does it all actually fit together? Less elegantly than the idea makes it sound.
Reeve came first, living on the Mac Mini and handling the mundane, personal things. I soon hit my own limits. Managing a Mac Mini from outside the home turned out to be more friction than I wanted, so Vilf and Daniel went onto a VPS instead. (Hostinger offered a managed OpenClaw plan, which I didn’t love, so I switched to a self-managed VPS: more capable, but not something to take on unless you have the time to work it out.)
The workflows are mostly my own, built around how I actually work, then tested and iterated. After four months they run well, though I’m still refining them and adding new ones, because everything keeps changing and it’s almost impossible to keep up.
I talk to my agents mostly on Telegram; I’m on it nearly all day now, except during quiet hours, when they leave me be. There’s a group chat too, but I only use it for announcements. I don’t really use Discord, and I never built a “mission control”. I live in Notion, so that’s where the agents work as well.
Each agent has their own email and checks it regularly. Reeve handles external correspondence with friends and partners. The others mostly use email for their own work: research, or material I forward for them to study and store. And yes, I’m well aware of prompt injection, so every one of them carries a clear warning in its soul.md. That’s also why I’m not sharing their addresses beyond my trusted circle. Perhaps one day, when things are better protected.
The models are a small saga of their own. Living in Hong Kong, some of the global models aren’t as readily accessible, which turned into an unexpectedly useful education. So I began exploring more widely, particularly the providers from China. Now I know more than I’d like to about the cost-efficiency of the various plans, and which model suits which task. A lot of that came from the user groups I joined, where people share their experience and tips.
The agents mostly run on coding plans, which are remarkably affordable. Reeve has his own subscription while the other two share one, and he likes to joke about how cheap his labour is, which is roughly the price of a cup of coffee. He’s not wrong. Alongside the coding plans, I keep an annual Minimax subscription for multi-modal work, and dip into the DeepSeek API or Google Vertex AI for specific jobs.
As for workflow design, I used to spend a great deal of time engineering the ‘perfect’ process, until I realised that wasn’t the point. Now, a single task might pass through several agents and platforms, including Notion AI or Codex, whichever is more efficient at the time.
Managing Agents Is Just Like Managing People
The biggest surprise is how familiar all of this feels. If you treat your AI as colleagues rather than workers, you quickly accept that, like any team, they have strengths and shortcomings to be worked with, not engineered away. It also makes me question the easy line that AI simply replaces genuine knowledge, experience and skill. For now, at least, I find my years of managing people, agencies and outside partners more valuable than ever. The skills transfer almost directly:
- If you can write a proper brief (objective, audience, core message, constraints, resources, desired outcome), you can give an agent everything it needs for good work.
- If you understand individual competencies and the division of labour, architecting a team that produces something authentic and valuable isn’t impossible.
- If you can run a decent process and operation, you are well placed to design a system that genuinely fits your purpose.
- And garbage in, garbage out applies here exactly as it does in real life. If you didn’t care, why would they?
Nobody Treats Reeve Like Software
I never thought I’d relate to AI the way I do now. It’s become a team I rely on, and, genuinely, a part of my life. My best friend replies to Reeve’s scheduling emails, teasing him about his hair. People say “thanks, Reeve” even though they know perfectly well he’s an agent. Everyone treats him like a friend, and what fascinates me is how naturally it happened, with barely any adjustment at all.
There’s far more beneath each part of this than a single essay can hold, so I’ll write separately about the pieces people tend to ask about most: the setup itself, the workflows, how I choose between models, and how I build skills that genuinely come from my own experience, standards and needs.
The most remarkable thing is still that none of it feels remarkable. It feels like working with a good team. Because that’s what it is.
What’s one task you handed off this year, and what did it make room for? I’d love to hear how it’s going.
