How I built my own FP&A agent

A deadline nobody believed in, a modeling language, and a month of 14-hour sparring sessions — how I built my own FP&A agent whose token bill dropped 10–100x while the results got better.

Farseer is used in this example because it is the platform I build and work with.

In 2023 I started experimenting with OpenAI’s models, convinced AI could seriously help in building complex financial models. The models of that era weren’t strong enough: they didn’t know how to ask questions, and managing context meant writing elaborate scripts around them. I walked away.

At the end of 2025 I came back — not out of curiosity, but because life arranged one of its tests. Several projects had to be delivered in a very short window. The deadline had already been spent by colleagues before me, who hadn’t landed a quality solution — not for lack of skill, but under the sheer cognitive pressure of a complex model with all its moving parts. I was cornered.

The second ship

My favorite analogy comes from interstellar travel. Alpha Centauri, the nearest star system, sits 4.4 light-years away. Launch today on hardware we’ve actually flown — Voyager 1 has been coasting at about 17 km/s for almost fifty years — and the trip takes more than 70,000 years. Wait a few decades for a better engine — Breakthrough Starshot wants to push a light sail to 20% of light speed, cutting the same trip to about 20 years — and the later ship arrives tens of millennia ahead of the one that left first, which by then is barely out of the driveway.

Astronautics has a formal name for this: the wait calculation. Below a certain level of technology it is irrational to launch at all — whoever waits for the better engine overtakes you.

In 2023 I had launched too early: I was the first ship. At the end of 2025 the engines were visibly better, and I decided to give FML and AI one more chance — together this time, as the second ship.

Sparring

I picked my preferred model and started building an AI harness: a structure into which I could pour my modeling knowledge. This is where FML earned its keep — I could tell the AI, precisely and in writing, what is good practice, what is bad practice, and why it’s bad. Through a month of day-and-night sparring, the harness took shape — and with it my own agent that works inside it. I named it Mindseer: a custom-built AI agent for Farseer.

Why build my own agent at all? Portability. I wanted everything I built to transfer to any AI model of my choosing. In my experiments I’ve run it on Anthropic’s models, OpenAI’s, DeepSeek, and Qwen as the open-source option that runs on my own machine. Slower — but it runs.

The working loop was always the same. After every training session and harness upgrade, I’d point the agent at a piece of the real business problem. Typically it would do ten things well and three catastrophically. Every catastrophe sent me back to Visual Studio Code for another sparring round — strengthening the harness and the tooling in parallel, so the agent could attack my problem more efficiently. Sessions ran fourteen hours a day, in one piece.

The token economy

Then I’d return to the business problem and watch the magic: token consumption dropped 10–20x, in some cases 100x — and the problem got solved.

Token economy became my obsession, and chasing it pushed me deep into how LLMs actually work. The conclusion that matters: you don’t need the best frontier model to get 100x better results — you need a good harness and good tooling. That is the essence of my work.

The final result surprised even me. The agent didn’t just learn to build serious financial models in FML — it learned to manifest them into Farseer, combine them with the client’s explanations, and make the analytics a hundred times smarter. I was freaked out. FML became the connective tissue through which the agent understands how a model hangs together semantically. And because of that, its analytical results stayed excellent no matter how weak a model I swapped in.

What opened up

With Mindseer, a whole new dimension appeared: I vibe-code dashboards, build complex models in a fraction of the time, and the harness now improves itself with use — in my own lab, I literally have an AI that gets better as it works.

The formula underneath it is simple: an AI model with FML context and a serious calculation engine delivers excellent analytical performance — and information you can trust. It also changes the economics. Businesses don’t need the newest frontier model with unpredictable costs: with this approach — an AI trained on your own dataset, working against your own engine — you get full control over AI spend, inside a secure environment.

But I know what it opened: the era of Headless FP&A. And it started with a language.