Linara Bozieva spent 11 years at eBay’s analytics division before being laid off in 2024. Rather than re-enter a job market with more candidates than openings, she launched Ravenopous, a growth marketing agency in San Jose, and staffed it entirely with AI agents. According to Business Insider, she now runs 27 custom agents across a three-layer architecture that handles end-to-end marketing for five clients. Her total AI subscription cost: under $1,000 per month.

The story has been picked up by outlets from Financial Express to Digit, mostly as a “future of work” narrative. The more interesting story is in the operational details.

The Architecture

Bozieva’s system runs on three layers, according to Business Insider. The directive layer defines each agent’s identity, knowledge base, and operating rules. The orchestration layer has six agents that function as the planning brain: market researcher, data analyst, creative director, finance, legal, and a routing orchestrator. The execution layer has 18 agents split into three groups: three building technical groundwork, ten driving traffic and awareness, and five converting traffic into revenue.

She originally built the system on Google’s Antigravity platform, then migrated to Claude Code after hitting Gemini Pro token limits. She pays for Claude Code, Codex, ChatGPT, and specialized tools like HeyGen and ElevenLabs, plus API access to multiple models. Claude Code serves as the primary orchestration environment, with APIs connecting to other models for specific tasks.

The architecture is not complex. It is a directed graph of specialized agents with clear handoff protocols, built using markdown files, skills, and scripts. Bozieva told Business Insider she described the system in plain language, had the AI produce the agent files, then tweaked as needed. Her eBay analytics background was what made the architecture viable: she understood data workflows well enough to structure the system, even without a marketing background.

The Economics

The cost structure reveals something about where AI-native agencies sit in 2026. Under $1,000 per month covers Claude Code, Codex, ChatGPT, HeyGen, ElevenLabs, and API access. That is the equivalent of a single junior marketing hire’s lunch budget in San Jose.

Once a client is fully onboarded, Bozieva spends roughly two hours per week overseeing their account. At five clients, that is ten hours of weekly oversight. She estimates she could manage 20 to 25 clients before needing to hire, and when she does hire, she plans to bring on operators who oversee agents, not traditional marketers who execute work manually.

The implied unit economics: if each client pays even a modest monthly retainer, the margin on sub-$1,000 in tooling costs is enormous. The constraint is not cost. It is Bozieva’s time for the parts agents cannot do.

The Hard Limits

Bozieva is specific about what she still handles personally. Client calls require reading the room, identifying where a client seems uncertain, and making judgment calls that AI cannot make from a transcript alone. Strategic decisions in data-sparse situations, where the agents propose multiple paths and there is not enough incoming signal to choose between them, still require human judgment.

She also flags a domain knowledge constraint that gets underestimated in “anyone can build agents” narratives: she could not build a healthcare agent system because she lacks the medical expertise to know which guidelines to include, which agents to create, or how to update the system over time. The agents execute; the architecture requires genuine domain expertise to design.

This is the pattern that keeps showing up. The agents compress execution time. They do not compress the expertise required to tell them what to do correctly.

The Broader Signal

Bozieva is not the first solo operator running a business on AI agents. Every week brings another profile in Business Insider, Forbes, or a podcast. The pattern is consistent enough to constitute a trend.

What makes this case worth noting is the specificity. Twenty-seven agents. Three layers. Named roles. A migration path from Antigravity to Claude Code driven by real token-limit friction. A cost figure under $1,000. A client capacity estimate of 20-25 per operator. These are operational numbers, not vibes.

The question is whether this model scales beyond individual operators into actual firms. Bozieva’s plan to hire agent-overseers instead of marketers suggests she thinks it does. The test will be whether a second operator, without her 11 years of eBay analytics experience, can step into the same architecture and produce the same quality output.

If yes, the agency model changes permanently. If no, what Bozieva has built is a highly efficient one-person business, which is valuable on its own terms but does not generalize the way the “AI workforce” narrative implies.

Fourteen client profiles trained. Five active clients. Under $1,000 in monthly costs. Two hours of oversight per client per week. Those are the numbers. What they become at 25 clients with a second operator is the experiment that matters next.