Spent the last five months building the FDE function at PostHog from zero. It's been weird and rewarding.
This is a sort of internet checkpoint for me. I want to log some lessons from building FDE and exploring it as a truly AI-native team, hoping it helps others.
The people are the best part: customers end up solving problems I had whilst I solve theirs, colleagues nudge me back to a better reality when something starts drifting. It gets so people make the work feel like a personal project. That's great until you stop consolidating, and you risk pushing too hard, so you drop things that mattered, then spend weeks chasing something with nothing to show for it, but the people keep pulling me back to the happy path every time. My lead, Simon Fisher, most of all. He gave FDE room to grow before it had proof.
On the FDE name, things won’t stay fixed in the industry so anchoring on a label is not my style. I see the role as a response to a shift in the industry, but what matters is the question. Things and people are flexing all the way down, and staying ahead of that so other teams can focus on what they already do best is what I’m interested in. The role and its offshoots are artifacts of exploring that.
Some stuff I built (beyond the classic admin stuff)
Deliverables split by audience: humans and agents. We should acknowledge this split intentionally if we’re to explore it effectively.
The FDE vault. Jon Lu suggested I start working out of an Obsidian vault and after trying it something was unlocked in my brain. It gave us the idea to have a customer brain repo per account, so the work is transparent by default and customers and colleagues get the benefit of it without waiting on me. It’s like my replacement, only it lacks my swag so I’m still needed to spark joy.
My best work happens when I'm not working. My thinking goes overdrive when I’m walking or taking a shower, so I built ways to tap into those moments such as working with my voice, and the only ceiling is to keep things compliant.
Doing, capturing, tracking, aggregating, reporting and improving the work is becoming a single action in the system I'm building. Exploring just how that’s possible self-drives me. Think of shipping a PR while doing the dishes.
I've avoided frontier lab APIs so much that I built whole systems on my existing subscription within AUP, including automated evals and ratchets. One of my main interests is using AI to obsolete problems and scale down on AI itself. It allows us to use AI for the truly novel things, it saves money and I get to feel like a maverick.
Shower thoughts from building with LLMs
AI-native goes deeper than the buzz. Most products marketed as such are AI-accelerated versions of human-native work. Native means the tool and the person improve each other. For instance, I sorted my AI conversations into folders by domain, then noticed how my thinking had reorganized to match that structure, I was able to make better connections, and that gave me better ideas to improve the system, then the improved system changed me again. If only the software gets faster, you built an automation and called it AI-native.
Builders can create novel architecture, or improve on existing paradigms, from first principles now. Anyone using AI mindfully and challenging themselves constantly can do this. Danilo Campos said it best in one of our LLM seances via Slack: the work is groping in the dark. I will keep that quote to heart. Someone poking around clumsily in a safe LLM environment is likely to find something novel in a way someone carrying pre-existing playbooks may not.
Traces are lessons, not errors. That's helped both as a personal practice and as a way to build with LLMs.
Correctness gets conflated with delivery, and delivery is mostly opinion. Deliberation is expensive because most of it is scraping opinion out of arguments. I’m interested in building such that all preference is serviceable without affecting the system’s core.
Asking better questions matters more than any particular answer. The same truth can be said a thousand ways, and you only know you got it if you understand the question.
Some customer lessons
Defaulting to honesty, even when it means admitting I messed up, is worth more now than ever. It's made everything easier with customers and colleagues.
People want to know another human has their best interest. They don't care how much AI you use if you show them you're on their side. This also teaches you a lot about using AI better.
Customers have told me to pass them prompts or skills instead of deliberating in Slack. Some run agents hard enough that their agents talk to each other. Others have AI versions of themselves handling comms. It’s time to meet users where they are.
Software is definitely changing, but people are changing even more. I think understanding human psychology will matter more for AI-native builders than shipping features. We're becoming guides, for the LLM, for people, for the space between them and how to explore it ethically and delightfully.