365 days of FDE

I’ve finished a full cycle around the sun as an FDE. Before, I was a quant, engineer, product manager, and an entrepreneur. How different was FDE compared to running a startup? Did I end up enjoying the work? What exactly do FDEs do? First,

Why not start up again?

I did, kind of, start up again.

After automating a lot of workflows in Adaface and moving to an advisory role, I wanted to find the next problem statement. One vision that stood out was: use AI to automate/augment portions of our work that we don’t love (admin, HR, political, etc. purely modern work bottlenecks). I imagined a world where we completely automate 97% of every knowledge worker’s “routine work” so everyone spends time on a handful of things that they care about (and find fun). We narrowed to browser work, and spent a couple of months building an automation framework, likely ending up with YC.

But we realized the clarity of the mission is not strong enough. We are finding users for a solution rather than creating a solution for the user’s problem.

I had two options. Do what I did with Adaface: push and push till we find one real user and then scale. Or, second: be an FDE, stay close to users, solve real problems for large enough users and deliver impact today. And while doing so, observe the patterns.

So, what did we actually achieve?

We had end-to-end ownership of solving users’ problems: customizing models, creating agent primitives, designing multi-agent architectures, building evals and deploying agentic workflows in air-gapped environments. Couple examples:

We created long-running cyber agents that find vulnerabilities to defend critical citizen infrastructure 24/7. These agents are durable, run for hours at a stretch, token-efficient and run in secure sandboxes.

Another involved designing a custom model for understanding architecture blueprints better than competitor trillion-sized models at a fraction of the size/cost. We use these models to perform multi-hop reasoning on extremely large images and find safety gaps to ensure critical infrastructure is built safely.

I also had the joy of being part of the team growing Singapore FDE, scaling it to Korea and Australia.

Shipping projects like these at this pace, was simply not possible pre-AI.

How did we do it?

Agents :D. But beyond agents:

First, deeply understanding the domain enough to know what the problem actually is.

Then borrowing from traditional software engineering: designing multiple solutions, understanding tradeoffs, creating prototypes to validate those tradeoffs, converting it into an MVP, then putting it in users’ hands in a staging environment, seeing where the solutions fail and hardening it for production.

Then there are steps that are specific to AI engineering: designing golden datasets (sometimes real and sometimes synthetic), creating multi-agent architectures, creating loops that customize the architecture, and using those loops to hillclimb on the datasets.

Then there are steps that are specific to consulting: How do you communicate the solution to users who might not have adopted AI in the org so far? How do you promise timelines to model customization that is technically a research project?

Lastly, there are steps specific to being an entrepreneur: How can we deliver this project today and not next quarter? And if we were to build the same agent again, how can we do it in 1/10th the time? Additionally, working with other GTM teams in pre-sales and renewals.

To achieve all of this, we map a small team of FDEs (often not more than 2) with complete responsibility for a single problem statement. It runs like a mini-startup with a single user and a clear mission.

Did I care about the team? About the impact I created?

After a year, I can say, it was a strong yes.

I found myself caring as if it’s my own startup. Over the year, thanks to agents, I got to spend more time with my colleagues than I would have expected. I made some real connections, and I am glad I was able to get that sense of belonging quite fast.

It was the same about the impact. I found myself often taking pride in the end result of what the agents were automating.

This brings me to the last question I want to talk about.

Do I miss being a (old-school) developer?

Yes. Sometimes. Within a single year, I went from using coding agents for minor features to building, reviewing and monitoring e2e features with agents. I still use my terminal, Vim, and editors, but I also built my own meta harness (which gets replaced with a new harness every few weeks). Along with dot-files, I now have dot-skills. It got nostalgic too soon. But what keeps me excited is this:

A lot of software engineers like finding problems, being challenged by those problems, solving them, and then sharing the solutions with others. Now that software is solved (more or less), software engineers are coming to solve other problems in the world. Engineers becoming FDEs is one proof of this. And I find this exciting.

2025 was the year of agents. 2026 is a year of multi-agents (swarms?). 2027 is going to be the year of factories? outcomes? Vague clues. But it will be exciting to find out.

There’s a lot of good to be done.

built with btw btw logo