What an AI agent builder actually does all day
An AI agent is not a chatbot with a nicer personality. It is a piece of software that has been given a job, a set of tools, and the judgement to know when to use them. Our sales agent reads an enquiry, checks the calendar, answers the question that was actually asked, and books the call. Our operations agent reads the night’s work, finds what broke, and fixes it before anyone wakes up.
Building one means deciding what the agent is allowed to do, what it must never do, and how anyone will know the difference. Most of the job is not writing prompts. It is plumbing and proof. The plumbing is connecting the agent to the systems where the work really lives, the CRM, the website, the database, the calendar, because an agent that cannot read the calendar cannot book anything. The proof is the harder half, and it is where most people give up.
Why proof is the whole job
Anyone can get an AI to produce something that looks right. The skill worth hiring is being able to tell whether it is right. We have a rule here that you will hear constantly: a green test is not a fix, and a check written from the same assumption as the code can never fail.
That sounds abstract until it costs you a day. It is the difference between “the page builds” and “a person on a phone can actually read it”. Between “the API returned 200” and “the customer got their answer”. You will learn to write the check that could fail, watch it fail on a deliberately broken version, and only then trust it when it passes. That order matters more than anything else you will learn here. Most people who arrive have never been taught it, and every one of them says it is the thing that changed how they work.
The kind of person who does well here
The people who thrive are curious and slightly stubborn. They read the actual file instead of guessing what is in it. They notice when two numbers on the same screen disagree. They are happy to say “I do not know yet” and then go and find out, and they are not precious about being corrected in front of other people.
The people who struggle are the ones who want a spec handed to them in full. We will give you a clear objective, the sources to read, the boundary of what not to touch, and the shape of what done looks like. What we will not do is tell you every keystroke, because working out the middle is the part that teaches you anything. If that gap excites you, you will enjoy this a great deal. If it worries you, this is probably not the right programme, and we would rather you knew that now.
What you will build in three months
In your first fortnight you will ship something small and live. It might be a single agent skill, a test rig, or a fix to an agent that is answering a question badly. It will be real, and you will walk it yourself as the customer before anyone else sees it.
By the end of the first month you will own a piece of a working agent, with the tests that guard it. By month three, people who have done well have built an agent from a blank folder to a live deployment, including those tests and the documentation that lets someone else run it without asking you questions. That is a genuine portfolio piece and it is yours to show whether you stay with us or not. It is also, in our experience, more than most people build in a year of a junior job, because nothing here sits in a backlog waiting for approval.
Where the work comes from
We are an AI company, not an agency that bolts AI onto marketing. We build a business its whole AI team: the website, the agents that work it, the CRM, and one dashboard to run it all from. That means the agents you build have a job to do and a customer who notices when they do it badly, which is a much better teacher than any course.
We have deployed 500+ AI agents and built 300+ AI systems, and we have been building with AI since 2019. The clients range from a martial arts school to a medical device consultancy to an M&A adviser, so the problems change shape constantly and you cannot solve them by pattern matching. You will not be working on toy problems, and you will not be given a sandbox version of the work while someone senior does the real thing.
How the training actually works
You will be paired with someone senior and given jobs with a clear brief: the objective, the sources to read, what not to touch, the output we expect, and how we will know it is finished. That structure is deliberate. Vague instructions are the main reason junior work goes wrong, so we removed them.
You will get direct feedback, quickly, and it will be about the work rather than about you. We keep a written record of every mistake we have made and what it taught us, and you will read it early. It is genuinely humbling, and it is the fastest way we know to stop someone repeating a fault that has already cost us a week. You will add to it too, because you will find new ways to be wrong and that is expected rather than punished. The only thing we treat seriously is hiding it.
How to apply
Send us one email with a link to something you have built, one paragraph on what broke and how you worked it out, and the sentence “I want the AI agent builder role”. No cover letter, no CV design exercise, no portfolio site required. We care much more about the thing you made and how you talk about it than about how the application looks.
The paragraph about what broke is the part we read most closely. Anyone can show a working project. Far fewer people can describe the moment it did not work, what they assumed, and how they found out they were wrong. That is the whole job, so it is what we screen for. If your application is interesting we will send you a small practical task, and it is a problem we have already solved so we can compare your approach to ours rather than get free work out of you.
