Alice & Marketa: designing AI into a one-person business

TL;DR

I run a two-agent AI system in production: Alice, who helps with life (budget, fitness, events, a relationships CRM), and Marketa, who helps with my marketplace (listings, bookkeeping, marketing). They live on a small always-on machine at home, talk to me over chat, and do real work with real consequences every day.

Why I built this

Running a marketplace solo (see Mini Marketa) means the operational backlog is infinite and the budget for staff is zero. Every repeated task, from listing a toy to filing an expense to drafting a campaign, is a candidate for delegation. The question that shaped the whole system: what would I actually need in order to trust an AI with this? The answers looked a lot like good delegation: a clear brief, the right tools and no more, an escalation path, and a running cost that fits the value of the work.

The system

Two agents, deliberately separate. Marketa's brief is Mini Marketa: photograph-to-listing pipelines, expense capture into the accounting system, marketing content, ad-performance insights, order awareness. Alice's is my life: a budget that carries over month to month, fitness data sync, an 82-person celebrations CRM (birthdays, kids' names, gift history), event discovery that learns my taste, and a link vault. The separation scopes each agent's tools, data access, and personality to its domain, which shrinks the blast radius of any mistake.

Chat as the interface. Both agents live in my messaging app, organised by topic. There is no dashboard to remember to open. Operations happen where I already am: on my phone, in the moments between everything else.

Cost-aware model routing. Each kind of work runs on the cheapest model that does it well: a low-cost model for chat and scheduled jobs, a small vision model for reading photos and receipts, and heavy generation handed to a coding agent on a flat subscription (more on that below). Everything routes through one provider layer, so I can swap a model without rebuilding anything. Designing the routing rules is product work in itself: pricing each task against the value it returns.

An executor pattern for heavy work. The agents don't do everything themselves. For long-running jobs, they queue structured task files to a more powerful coding agent on the same machine, which executes and reports back. The chat agents stay cheap, responsive, and always available while the heavy lifting happens on a different cost model. Before trusting it I ran a full dress rehearsal of the queue end to end, because agents deserve release processes too.

AI in the loop, with me at the wheel. Anything customer-facing or financially meaningful comes to me as an approval, one tap in chat. The same pattern gates access to the full version of my AI Visibility case study on this site: a request arrives, Alice pings me, I approve or deny. Autonomy is a dial rather than a switch, and choosing where to set it per task is the core design decision of agentic products.

What it does, concretely

• New toy stock goes from photos to published, priced product listings with my review in the loop.
• Business expenses flow from receipts and emails into the accounting system, categorised.
• Marketing tasks (campaign drafts, content, performance summaries) get produced against the brand's locked style rules.
• I get told before birthdays, when a bill is anomalous, and which events this weekend I'd actually like.

What building this taught me about agentic products

Trust is designed, not accumulated. Users, me included, trust an agent when its failure modes are visibly bounded, not simply after watching it succeed. Scoped tools and approval gates did more for adoption than any capability improvement.

The economics are a first-class design surface. Model routing, executor offloading, and choosing what not to automate determined whether the system made sense at all. Agentic product design is unit-economics design.

Agents need management, not just prompts. Job descriptions, feedback that updates their instructions, dress rehearsals before new responsibilities: the operating rhythm of running these two agents looks a lot like being a lead PM for something very fast and very literal.

Add AI where it belongs, not where people were. I think good product management and design finds organic, natural ways to put AI in the loop, rather than replacing humans with it. Marketa and Alice absorb work that had no one to do it, and that framing shaped every design decision above.