This as-told-to essay is based on a conversation with Linara Bozieva, a 39-year-old founder of Ravenopous, based in San Jose. The following has been edited for length and clarity.
A little over two years ago, I was laid off from eBay after spending 11 years with the company. My family and I had recently moved from Switzerland to the US, and it was a hard time.
I looked at the job market and saw that many companies were laying off workers, and there seemed to be more candidates than openings. It felt worth it to start building something small, at least for a couple of years, to support my family and me.
A few months after my layoff in 2024, I launched a traditional marketing agency. I don’t have a marketing background, but I built a three-layer AI workflow with 27 custom AI agents that run an entire marketing strategy under my oversight.
My professional background helped me automate my business
When AI became mainstream, I used ChatGPT in my browser and tools like Midjourney for video stuff.
Then I started seeing how people were using AI in these closed-loop systems, where agents acted autonomously, and I wanted to build something like that.
I originally built my system on Google’s Antigravity platform because it was very user-friendly, but I eventually switched to Claude Code after running into token limits with Gemini Pro. I love Claude, but I have started hitting token limits there as well.
I worked in analytics at eBay, and that background helped when I was building the architecture and creating guidelines. At the same time, AI wrote a lot of the system itself. With all the markdown files, skills, agent files, and scripts, I told the model what I wanted it to do in plain language, the AI produced it, and then I tweaked as needed.
Here’s how my 27 AI agents function on my team
My automated system has three layers: directives, orchestration, and execution. The directives layer defines who the agents are, what they know, and how they operate. The orchestration layer decides which agents should do what and when.
There are six agents in the orchestration layer that act as the brain and think before any other agents act. Those agents are a market researcher agent, a data analyst agent, a creative director agent, a finance agent, a legal agent, and an orchestrator agent that routes tasks to the execution layer.
After the tasks are routed, three agents build the technical groundwork, 10 work on driving traffic and awareness, and the last five agents are responsible for converting traffic into revenue. The execution layer includes scripts that help guide repeatable tasks.
The entire setup costs me under $1,000 a month. I’m paying for Claude Code, Codex, and ChatGPT. Then I have a few specialized tools like HeyGen and ElevenLabs, and I’m paying for APIs because I’m also giving my Claude Code access to other models and tools through APIs.
The hardest part was knowing when the system actually worked
The system has now been trained on 14 clients in total. I currently have five clients, but I also tested the workflow on my own projects and did free work for friends to improve the agents.
The most challenging thing was knowing when the system was operational. Which is why I tested it on 14 client profiles. I wanted to make sure that the system fully understood my approach and frameworks and delivered high-quality results every time.
As a result, I now have Gemini and Claude generate my ad copy separately, then compare and combine the results to get something that feels more human. The final best copy is produced from their combined output.
I also added scripts in the execution layer that pull customer pain points from Reddit discussions to sharpen the strategy.
Read more from our Tiny Teams series
Social skills are what I’ve realized to be very valuable in the AI era
With my business, the AI agents run ads, analyze performance, improve creatives, and then give me daily reports. The client gets weekly reports, and we do broader monthly calls.
AI can process a transcript of a client call, but it cannot fully read the room or identify where the client seems most nervous. That judgment and client management still have to come from me.
I’m also heavily involved in the strategy when there isn’t enough incoming data, and my agent and I are at a crossroads about which path to choose. I continually check if what the agents propose really makes sense and is realistically doable as a strategy.
Also, AI can build you a system, but you need to understand the domain. For example, I wouldn’t be able to build a healthcare agent that treats colds and flu in children because I don’t have that knowledge and wouldn’t know which guidelines to include and which agents I would need, or how to better update the system as I go.
I think all repetitive tasks can and will be done by AI in the future. However, the expertise needed for specific departments or domains will still be with humans acting as VPs over AI architecture.
AI changed what I believe is possible for one person
I still have a few traditional marketing clients that I serve with my freelancers. Since we’ve been serving them before, we can’t say we aren’t doing that anymore. It’s a smaller, separate branch of my business.
For all the new clients I take on, I now provide only full-cycle marketing through the AI system. Once a client is fully onboarded, I spend about two hours a week overseeing their account.
If I focused only on client oversight, I think I could comfortably manage 20 to 25 clients on my own. Once I reach that number, I’ll only hire people like me, operators who oversee the agents and guide the strategy.
Building something on my own has been incredibly fulfilling and promising. It changed not just how I work, but what I believe one person can build.
Do you have a story to share about working with mostly AI agents? If so, please reach out to the reporter at aapplegate@businessinsider.com.