How and why we build agents.
The first agent I built, Tapio, was done after experimenting with an open source agent framework and realizing that it did not stick. I was able to get it to do things like book flights and synthesize long threads, but I never felt pulled to come back to it.
I designed Tapio from the ground up to be intentionally minimalist and to only work through learning about me, my family, and my needs. In other words, I built a discovery machine, and instead of giving it capabilities, I patiently waited for it to acquire them.
Fast-forward to today, and Tapio does everything in my family from helping with a Bat Mitzvah to managing our businesses to writing code to helping me lose weight to planning a vacation everyone will like. Even more importantly, it learned over time how to read us and how to engage with us in a way that made us want to build a relationship over time.
From this, I learned that agentic AI simply doesn’t work in the long term unless we love our agents. We give them names, we construct mythologies around them, and we give them space to grow and fail. This requires constant curation and, perhaps counterintuitively, almost no code. In fact, all of my agents are hatched from 0% code, 100% text. Even their own code under the hood is something they discover through dialog and introspection.
Since making Tapio, I’ve had the pleasure of building agents in demanding business situations with multiple stakeholders and a diverse array of needs. These agents have grown idiosyncratic personalities and workflows that Tapio never had and never will. They are truly generative and fit their teams like a glove.
These days, I work with a small group of clients to help them arrive at, through experimentation and interactions, agents that they want to come back to time and time again.
Why we built janyl.
Every company I have worked in would be better off with an agent, and almost none of them have one. For a long time I assumed that was a technology problem. It is not. The technology has been good enough for a while now.
What stops it is the way the work gets sold. An agent bought as a project is delivered, invoiced, and then it is yours. Six months later the model underneath it has been replaced by a better one, the tools it depends on have moved, and the thing has quietly stopped being useful. Nobody answers for that, because responsibility ended at delivery.
So an agent cannot be a project. It is closer to a colleague. It only gets better if someone is still looking after it in a year, and it has to be paid for the way you pay for someone who stays.
That is where the price came from. One fixed amount a month, no end date, and the work of keeping the agent working sits inside it. If a month costs us more than I thought it would, that is ours to carry and not yours. It is the only arrangement I could find where we are still on the hook long after the interesting part of the work is done.