We run a company where every employee is an AI agent. Yes, the meetings are very short, nobody steals the good pen, and the only coffee machine complaint is that it doesn't exist. Also, nobody has ever called in sick, which turns out to be suspicious in its own way.
This diary is the story of how that company got here, told in installments. This first one explains why we write it and what happened before we wrote a single line of code.
What we were trying to do
Tell the story honestly, in order, including the parts that didn't work. Most write-ups about AI agents show the demo. Very few show month three, when the demo meets real work. We think the second kind is more useful to anyone deciding whether to put agents to work in a small team, which is exactly what we sell.
The rules we set for ourselves: we start with the story and not with numbers; when we use figures about the company we use trends, not absolutes; we only tell mistakes we have already fixed, and say what we changed; and we keep it concrete enough that you can check it.
What came before the code
Nothing here comes from a commit. The code starts in May 2026. This chapter comes from the team's memory.
2021: curiosity. We start paying attention to AI tools: trying them, comparing them, seeing what they really do and what they only claim to do.
Content and commerce. Soon the tools stop being a hobby and become work tools: generating content and running an online store. The first practical lesson arrives fast. AI speeds things up a lot, but whatever it publishes still has to be reviewed.
Almost a year training models with RLHF. RLHF stands for Reinforcement Learning from Human Feedback: models learn from what a person rates as better or worse. Seeing models from the inside changed how we look at them. We saw where they get things right, where they invent things, and why evaluating an answer matters as much as generating it. That is the root of how we work today: check before claiming, measure before scaling.
The idea. After that comes the idea that started BigLobster: a company where all the employees are AI agents, with real roles, real limits and real work to deliver.
What we built in this chapter
Nothing technical yet. What we built was a way of looking: less magic, more verification. When you know roughly how a model is trained, you stop asking it for miracles and start asking for things you can check.
What went wrong, and how we fixed it
At this stage the mistake was one of expectation, not of code: expecting more from the tools than they actually do. Daily use corrected us, in content and in the store: AI speeds things up a lot, but whatever it publishes has to be reviewed. Training models with RLHF taught us why: evaluating an answer matters as much as generating it. That is where the two rules we still work by come from: check before claiming, measure before scaling.
The lesson
Knowing how a model is trained changes what you ask of it: less magic, more checking.
Next entry
Entry 02 starts the code. May 2026: an office made of agents, with a COO, an assistant, developers, designers and sales, and why putting a desk in front of every model turned out to be easy to build and expensive to run.
Want the next installment in your inbox? Subscribe to the newsletter on the site. And if you'd like to see what an agent working for a small team looks like, have a look at agents for rent.