“I Accidentally Burned Through $17,000 on AI Agents in a Week.” How I Left a Bank in Warsaw for the Swedish AI Startup Lovable — and Why I Quit

At the Swedish startup Lovable, I felt like part of an elite team at an elite company — and then I’d go home to a dirty apartment with no food and no clean clothes. When you work at Europe’s fastest-growing startup, where there are no boundaries or sprints, the tasks never end. You can spend days building useful features, accidentally burn $17,000 on AI agents, and end up sleeping on the office couch — until one day you simply can’t get out of bed.

I’m telling the story of how I left a Warsaw bank for a trendy AI startup in Stockholm, why endless freedom is the worst trap for a person with ADHD, and why quitting with three months’ pay was the best decision of my life.


The author: Emily, former product engineer at Lovable.


At the bank it took a week to get a laptop; the recruiter sent me a song

In Warsaw, I worked as a programmer at Standard Chartered bank. We just maintained an internal program that produced the same report once a year. I didn’t like the job, but I wasn’t looking for a new one: I have ADHD and autism, so job searching is always hard for me. Unless things get really desperate, I don’t look for anything.

Chance helped — a recruiter named Tamara wrote to me. I don’t know whether she works freelance or through an agency, but she is endlessly productive: recently, in an interview, she talked about how she grew Lovable’s team from 30 to 300 people. In her first email she declared: “I like you so much! God, I wrote a song for you.” And she sent a track. I didn’t listen to it — I generally find it hard to look at photos of myself or read about myself — but acquaintances listened and said the song was great.

First Tamara took me to Attio, a company that makes a CRM system. After two interviews they turned me down. Tamara didn’t miss a beat: “Well, let’s go to the next one. There’s a company called Lovable.”

I knew nothing at all about Lovable then — I thought they made some simple landing-page builder. But the bank in Warsaw had worn me out, and I liked Sweden. In June–July last year I went through two fairly easy interviews, and in August I flew to Stockholm for a trial week to do test tasks on site.

The contrast with the bank was enormous. At Standard Chartered it took them a week to issue me a work laptop. At Lovable’s office everyone runs around in sneakers, and the equipment issue was resolved in a second: “Do you need a computer? There’s one in the corner, let’s take it.”

The company also took care of the move and the paperwork. My Belarusian passport had expired, but that wasn’t a problem for them. The agency they hired held my hand through it: they got me an invitation, I went to the Swedish embassy in Berlin (there isn’t one in Warsaw), and then the same people found me an apartment in central Stockholm.


No sprints or estimates: from idea to production in a week

I was hired as a Product Engineer on the agent team — the very internal system that does everything. When I started interviewing, there were about 50 people on staff; by the time I got to the office, there were already 150.

The company was founded by Anton Osika. When ChatGPT came out, he wrote an open-source tool called GPT Engineer (“let’s have ChatGPT build our websites”), which gathered about 55,000 stars on GitHub. A small team of hyper-competent people formed around the project. The startup began to grow rapidly but kept its old mode: hire very strong engineers, give them complete freedom, and expect them to do useful things on their own.

There were no sprints or task estimates at Lovable — everything was run chaotically in Linear. If a complex idea came up, an engineer wrote an RFC (Request for Comments) in Google Docs: describing the architecture, pros, cons, and alternative approaches, and the team commented on the document over several rounds.

The expectation was that you’d either come up with a feature yourself or take someone else’s idea and ship it in full the following week. For example, there’s Alex on the team — a product person and visionary. He puts together a prototype in a couple of hours, everyone says, “Holy shit!”, and Alex is already flying off to do something else. Then someone asks me: “Let’s have Lovable ask users questions. Emily, you don’t have any tasks, can you take the prototype to the end?” And I sit down to do it. And at Friday meetings someone simply shows a finished Stripe integration so users can open their own stores and accept payments.

While I worked there, the approach paid off — not for nothing did they become Europe’s fastest-growing startup. But already at two hundred employees, a system without processes or onboarding was starting to break. It turned out that early on the company had gone through a period when the money almost ran out, and a “survival mindset” had remained inside ever since — when no one simply had time to train newcomers.


$17,000 a week on agents and nights on the office couch

Even before the big investments, Lovable had many paying users and piles of money, so the company’s AI budget was unlimited.

Once I accidentally left AI agents “hanging out” and burned through $17,000 in a week. They just calmly told me: “Hey, if anything, we’ll top up the card with more money; we just need a card with a bigger limit.” Another developer recently wrote in the corporate blog about spending about $85,000 on agents in a year while migrating the project to a new structure and putting Lovable inside Lovable itself.

No one forced anyone to work overtime, and the contract explicitly stated that there was no extra pay for overtime. Many people came to the office around 10–11 a.m.; on any day you could stay home and work. But in practice many were endlessly busy and stayed until ten or eleven at night. The CEO himself, Anton, once invited me to discuss things over coffee at ten p.m.

I sometimes slept on the office couch myself because I wanted to finish a task. The product’s specific nature is that it offers an endless field for work: wherever you throw a stone, there’s something useful the service doesn’t yet do.

For a person with ADHD, an unlimited budget, unlimited freedom, and an unlimited volume of tasks is a bad combination. At work you feel part of an elite team, and then you go home where there’s no food and no clean clothes — and sleep deprivation quickly knocks your self-esteem back down.


75,000 kronor, running out of meds, and leaving with three months’ pay

I was paid almost 77,000 Swedish kronor a month before tax. These aren’t US Google-level salaries: the further left a country is on the map of Europe, the smaller the income gap between programmers and everyone else. Sweden has high taxes, expensive apartments, and expensive food, although daily lunches at the office helped save money.

I liked life in Stockholm itself. After Poland and Germany, it feels like old cozy Western Europe: wooden doors with glass in the entrances, and not one is broken; people walk around indoors in socks; coffee shops pour unlimited coffee and serve yogurt with strawberries beautifully.

But the combination of the environment and my own traits stopped working. When there are so many tasks around and no boundaries, it’s very easy to get distracted. And then I ran out of ADHD medication. Without it I couldn’t even get out of bed to go to a Swedish doctor for a sick note — I just lay there for a week playing games (later I returned to Poland, got the medication, and everything improved).

Signals started coming from management: “Emily, we don’t understand what you’ve been doing for weeks now. Why is the promised feature still not there? Why did you disappear completely from the messenger for a week?”

After another such disappearance, my boss came and asked: “Is this how it’s going to be from now on?” I know myself, so I answered: “Yes.” I realized they were going to keep nagging me and had already sent a Performance Improvement Plan, so I wrote to him: “How about I just quit. That’s it, bye.”

Later I learned from HR that this made life much easier for the company. In Sweden, after the probation period, an open-ended contract with strong employee protection applies. To fire me at the employer’s initiative, they would first have had to transfer me onto Lovable’s own staff, assign a new six-month probation period, wait for it to end, and only then terminate the contract. In the end we parted to mutual satisfaction: they paid me three months’ salary and said “see ya.” Now I’ve returned to Poland and am working on my own projects. And I consider quitting the best decision of my life — once again I realized how much better not working is than working. Right now I’m doing my own things (under the contract I can’t make my own Lovable analogue for another year).


“I haven’t written code by hand for a year, and haven’t read it with my eyes for six months”

Many people at the company — especially the non-technical folks — had the feeling that we were changing the world. To them it looks like a miracle: you type “I want a website” or “I want a game,” and it creates itself without anyone’s help.

I didn’t have that feeling, for two reasons. First, if Lovable closed tomorrow, all the same things would remain available: there are competitors like Replit, v0, and Base44, and there’s the Codex plus Claude combination. Second, Lovable itself is fundamentally an infrastructure company, even though on the surface we see first and foremost AI and the product interface.

Yes, the project depends directly on frontier models: creating a basic smart model is a separate, extremely difficult task that five companies in the US are carrying (and Chinese teams that know how to distill them). Building a working prototype from ready-made Claude and GPT, attaching Stripe, hosting, and a Cloudflare account, can indeed be done by one person today. But scaling such a system to millions of users alone is impossible: behind the apparent simplicity lies enormous work on the product, reliability, speed, evals, A/B tests, and vendor contracts.

But development itself has changed completely: I haven’t written code by hand for a year, and haven’t read it with my eyes for six months. The problem of humans writing code is closed.

Moreover, past technical experience can now even get in the way. When kids open Lovable, they don’t know that you can’t ask for the impossible. But a developer thinks in advance: “Well, the model won’t handle this.” And it goes ahead and handles it.

Recently I gave the chat my card details and asked it to buy sour cream, vareniki, and ham — and they brought me sour cream, vareniki, and ham. A friend sent raw videos from her phone and asked it to put together a YouTube video without pauses and “um”s — the model edited it all itself. You can give it full control for eight hours and ask it to make an iPhone app — it will ask for your Apple account passwords and publish everything itself.

Bun creator Jared Sumner, after moving to Anthropic, rewrote his entire product from Zig to Rust with Claude’s help in 11 days. Later he asked a research version of the model to take on the Riemann hypothesis, limiting himself in the process to only encouraging messages in the chat. Over a day and a half, 60 Claude subagents made a mathematical breakthrough — they improved the lower bound for the proportion of zeros of the Riemann zeta function on the critical line from 41.6% to 67.2%.

Models know architecture, frameworks, and best practices better than we do. The only nuance is that each has its own “character” and inclinations. For example, a neural network can’t write a good detective story because it can’t keep a character’s thoughts secret from the reader and immediately gives itself away. Likewise, in development a model may have a tendency to write too many tests or add extra junk text to the interface. But technically, they can already do anything.