An estimated 36 person-years of work, built by two people and AI agents in two months

Where the value sits

Any company can start AI agents on building and recurring work today. Starting them is the easy part. What lets a company leave them running is judgment written down, the rules, the quality bars, who approves what and what it has learned. That written layer is what the company owns, because agents can be replaced and the models improve.

Flip Education ran on this. I directed in working sessions, live sessions where I set the direction and AI agents wrote the code, and they built the product that way. In the background, agents ran the recurring work, each with its own job. Under both sat the written rules and what we had learned. Those rules set what every agent could do and what its work had to pass, so both kinds of work held the same standards.

The challenge

A company built by two people has a simple arithmetic problem. My co-founder, a career educator, carried pedagogy. Everything else was mine, from engineering direction to go-to-market, operations, legal and support. Launching in that many markets meant building the product plus curriculum pages and content for every market, in each language. No two people do that by hand. So the question from the first day was which parts of the work a machine could carry, and how much trust each part deserved.

The organisation

I built the product in working sessions, live sessions where I directed and AI agents wrote the code. My direction set what got built, and the same written standards later held for the agents too.

When work repeated, I wrote each agent down, working with an AI agent. Each agent was one written entry with a role, a permission level and a definition of done, plus how it ran, on a schedule, live with me, or started when needed. So before its first run, one place said what each agent did and how far it could go. From then on the agents ran as their entries said, and the scheduled ones ran without either of us starting them. Automatic checks, some of them AI models, had to pass before work went on, and failed work stopped there, so a broken change never reached a customer unchecked. Agents watched the other agents, and one agent checked the claims in agent reports against the raw data before we read them. The two of us read reports that had been checked, never every run. A short written list of actions always needed a person, for example moving money, changing what customers see and emailing anyone outside the company. And when something went wrong, the fix became a rule or check that later work had to pass. The actions that could hurt the business stayed with a person, and the same mistake didn't come back.

What happened

The code history starts in February with a working first version of the app. 8 days after the first version, the code carried pages for every country we served.

What it would have cost

The estimate puts the work at about 36 person-years, meaning one person working for one year, times 36, about €1.5 million at Spanish pay. Pay was priced for Spain, Germany and the United States, and Spain gives the lowest figure, so it keeps the estimate cautious. It is an estimate, built from a standard software cost model for the code, assumed times per topic for the curriculum, published writing and translation speeds for the rest of the content, and published pay for each role in Spain with employer costs. It assumes a hired team in 2026 would also write code with AI help, so the software part was lowered for that. A hired team would have been about 20 people. The work would have taken them nearly two years. No hired team could have done it in two months, because even the most cautious version of the estimate needs most of a year for the software alone.

What made it possible

Two things, I would say. First, my 14 years in go-to-market and operations. I knew how to move a company forward the old, manual way, and that business sense is what let me direct the AI instead of starting at random. Second, luck in the timing. Agents had just become good enough, and they get better every week.

What I would keep

I'd keep the trust steps, every agent starting only allowed to look and earning more in writing, with the guards that went with them, tests first, spending caps, reviews before any guardrail came off. And I'd keep rules built into the tools themselves, so nothing moves on until the tests and checks for that piece of work have passed. That is what let agents run every day with nobody watching each run.

If your team builds this way

Don't build this way if you don't yet know the work well enough to judge it. If you can't tell a wrong result from a right one, agents repeat the wrong one every day. In a running company, move one area at a time and keep a person checking the results before you widen the change.