How to build your own AI employee

Everyone is talking about AI employees. Few people explain what one is made of. Without that, you end up with a chatbot you have to babysit all day.

Tomi, founder of SupaKid Studios, says last month he spent over $20,000 on AI subscriptions and tokens running multiple teams of AI employees. He is clear that most people do not need to spend anywhere near that. What he learned is how the pieces fit, and that part is cheap to copy.

The three parts

Every AI employee has the same three parts.

1. The model. The brain. It reasons and responds. Examples are the models from OpenAI, Anthropic and Google.
2. The harness. The software that runs the employee and manages its work. It decides what the model sees, keeps track of tasks and lets the model use tools. A model without a harness can only talk.
3. The tools. The apps and capabilities it uses to get the job done: a browser, your task board, email, code editors, a computer of its own.

If your AI employee disappoints you, check which part is weak. A smart model with no tools cannot do much. Good tools with a vague job go nowhere.

Step 1: Choose one job

Tomi's advice: start with one job you actually need help with. Not "run my business." Something like:

• Update the task board after every meeting.
• Research ten possible customers and summarize them.
• Draft replies to common support questions.

Pick something you can check in minutes. That makes it easy to see if the employee is any good.

Step 2: Give it context and memory

An employee who knows nothing about your company will make things up or ask you everything. Give it:

• Context: what the business does, who your customers are, what good work looks like.
• Memory: a place to keep notes about decisions and preferences, so it does not start from zero each day.
• Examples: two or three finished versions of the job.

Keep it relevant. A long dump of everything makes the output worse, not better.

Step 3: Give it only the access it needs

Access is where things go wrong. If the job is updating a task board, it needs the board. It does not need your bank, your private chats or your other clients' data. Add access one step at a time as you trust it. Remember you stay responsible for anything it does in your name.

Step 4: Review the result

Tomi's rule is to define the input and review the output. The loop should be fast but small. If the result is not what you expected, fix your instructions, not your tolerance. Keep a person approving anything that goes out to a customer.

How we do it at SupaKid Studios

Tomi's work spans code, content, a game, ads, an engineering studio and financial services infrastructure. He is building toward employees that keep working around the clock, with him directing and reviewing the work. For software, agents write the code, another agent tests it, and the results come back as a pull request for review. We call that an AI software factory when it is set up across a team.

Client example: SkipHold, a white-label AI communications platform, wanted their team building with AI. We connected them to AI development and software factories, agents that build, test and ship code. For 0G, we run engineering with agents. For small businesses, the job is often much simpler, like the bookkeeping agent we set up for Blockwise.

Once you have more than one employee, the question becomes how to organize them. Read the AI chief of staff guide.

What goes wrong

• A job that is too big. Split it.
• No review. Employees repeat their mistakes until you correct the input.
• Too much access, too early.
• Confusing a chatbot with an employee. A chat window is not a harness.
• Hiring five before one works. Get one reliable first. Then see how it fits into a one-person company.

What to do today

1. Write one job, in two sentences, with a clear finish line.
2. List the model, harness and tools you will use for it.
3. Give it the context and examples, nothing more.
4. Run it once and review the output.

SupaKid Studios can set up, run and manage agents for you, or coach you through doing it yourself with coaching. A 30-minute call is enough to pick the first job.