How I built VillaSlot in 48 hours with AI agents

▶️ Watch the full video

Timely’s technical support took four days to answer me. Instead of waiting longer, I built VillaSlot, a SaaS tailored to the way my villas operate. The first version was developed in 48 hours, including a 9-hour and 20-minute autonomous run overnight.

This did not come from one large prompt asking an AI to “build a SaaS.” It came from product framing, functional decomposition, and a quality loop strict enough to let agents work for hours without constant supervision.

The business problem

My villas use two booking models. Travelers book nights, as they do on Airbnb. Photographers, brands, and event organizers book hourly slots. Both models must share one calendar to prevent conflicts.

VillaSlot also had to manage deposits and balances, then route payments to the relevant villas’ Stripe accounts. The product combined several constraints in one flow:

• nightly bookings;
• hourly slots for shoots and events;
• a shared availability calendar;
• deposits and final balances;
• payments to external Stripe accounts.

VillaSlot replaces the Timely and Airbnb workflows I needed. It does not replace Airbnb’s marketplace or customer acquisition.

Framing before coding

I started with a brainstorming session with Codex to define the scope and product decisions. I then prepared the PRD, 20 user stories, their acceptance criteria, and a dependency graph.

The dependency graph mattered because some stories relied on authentication, while others depended on the calendar, payments, or villa management. Running everything in the wrong order would have created unnecessary rework and conflicts.

I also prepared 20 mockups, giving each agent a precise functional and visual target.

A 9-hour autonomous run

Once the groundwork was ready, I launched the agents overnight. They worked through the stories for 9 hours and 20 minutes. The goal was not to generate as much code as possible, but to finish every story with evidence.

Each story followed the same pipeline:

1. define the expected behavior;
2. write tests with a TDD approach;
3. implement the feature;
4. run a separate review;
5. execute unit and integration tests;
6. test the complete flow in Chrome;
7. check the interface visually;
8. automatically fix issues until validation.

The next morning, the first functional version matching the defined scope was ready. It was not final. Real usage produced about ten additional stories the following day.

Why QA matters more with agents

The longer the autonomous run, the stronger the QA must be. Without precise criteria, tests, and visual checks, an early mistake can affect hours of subsequent work.

The developer’s role changes. Less time may go into typing every feature, but more discipline is required to define contracts, alternative cases, and the evidence expected from each agent. Agent autonomy depends on the validation system.

Token cost versus human time

The project processed more than four billion tokens, with most of that volume coming from cache reads. This should not be confused with four billion new tokens billed at the standard input rate.

The useful comparison is the delivered result: 20 stories, tests, reviews, end-to-end flows, and corrections. Token usage looks excessive when viewed in isolation. It becomes easier to assess when compared with the human time required to build, test, and review the same scope.

The Killer SaaS method aims for that balance: enough preparation and QA to turn long agent runs into a working product rather than a pile of fragile code.

Explore VillaSlot and watch the complete workflow demonstration.

I share practical lessons about AI agents and software development every week in The Agentic Dev.