Just Better 
Building AI agents
since day 1.
The principles we use to build them.
What is an AI agent?
AI, ChatGPT, GenAI, LLM... whichever name you know it by, to most people they all mean the same thing: a brain that can answer almost any question.
But a brain alone can't do much. It can tell you how to write the email. It can't send it.
That's why AI companies connect this brain to your email, your calendar, your files, your CRM, your browser. They're giving the brain hands. That's an agent: AI that doesn't just answer, it acts...kind of like an employee.
We build custom agents for your company.
Here are the principles we use to build agents
Agents automate tasks.
An agent can do almost anything: book a meeting in your calendar, create a quotation, answer customers, chase a late invoice, prepare your weekly report... But agents are really hard to build, which is why you don't see them everywhere yet. We've been building them since day one. Here are the principles we use to build them.
Principle #1
A task has a start and an end
One of my clients is a manufacturer: customers write asking for a price, his team writes back with a quote. Say you want to automate that task. When would it start? When the request arrives by email, or when a human hands it the request?
- If it starts when the request arrives, the agent needs to monitor the inbox (access to your email) and track which requests were handled (a database).
- If it starts when a human hands it the request: much simpler, the agent only turns a price request into a quote.
Defining the start and the end has the biggest impact on the complexity of the system.
I know the start and the end...what next? If a human still does the task, it's because that human makes decisions.
Principle #2
Every task hides more decisions than you think
When we started, the CEO told me the task was simple. A request comes in. You check the price list. You send the price.
Then I sat with the person who actually does this. Half the requests were missing information, so she wrote back to ask. The measurements came in the wrong units, so she converted them. Some terms didn't match the price list, so she looked them up.
The CEO described one decision. She was making twenty.
Seems obvious, but every single project I've started suffered from this. So map every decision and every exception. And to do that, ask the person who does the task, not the person who manages it.
Principle #3
Test agents using the scientific method
Ok, you mapped the decisions. Can you trust an agent with all of them?
Some uncomfortable math. If an agent makes 20 decisions, and makes the correct one 95% of the time, you get the right output one time out of 3.
You can improve results: fewer decisions, a human checking the output, verification steps. But it all comes from one principle: agents have to be tested using the scientific method.