I’ve spent most of my career in R&D. The nature of that work is that it generates questions faster than it answers them. You solve one problem and three more come into focus. You finish a project and immediately see the better version you’d build if you started today. The ideas don’t stop accumulating — they just sit there, in a mental queue that never gets shorter.
That’s not a complaint. It’s mostly what I like about the work. But it does mean that for years I’ve had a backlog of questions I genuinely wanted to answer and no realistic way to get to most of them. There’s only so much one person can run in parallel. At some point you make peace with the fact that most of the questions will just stay questions.
That changed when AI stopped answering questions and started doing work.
The shift that mattered
The chatbot phase didn’t move me. Question-answering systems are useful the way a good search engine is useful — they save time, they surface things you might have missed, but they don’t change what you’re fundamentally able to do. I wasn’t an early adopter. I spent that period watching, reading, trying to understand what was actually happening under the surface before I built anything.
The shift to agents was different. A system that can take a task, break it down, use tools, and produce real output — that’s not a better search engine, that’s a different category of thing entirely.
I set up my first agent and started to understand what was actually possible. And I started thinking about what it would mean to have not one agent but a coordinated team of them — specialized, persistent, able to run work in parallel. Something like a small lab. Agent researchers and caretakers that could finally start working through the backlog.
The centaur
Around that time I came across a framing from the chess world.
After Garry Kasparov lost to IBM’s Deep Blue in 1997, he didn’t retreat from the game — he invented a new format. In Advanced Chess, later called Centaur Chess, humans and computers played together as a team. The freestyle tournaments that followed, held from 2005 to 2008, produced a consistent result: centaur teams outperformed both the best human players and the strongest AI opponents playing alone. The human provided direction and judgment; the machine provided depth and calculation. Neither alone was as strong as the combination.
What struck me most wasn’t the result at the top — it was what it implied everywhere else. You didn’t need to be Kasparov for the combination to work. Average players paired with average machines were beating grandmasters. The value wasn’t in either partner’s individual strength. It was in the teaming.
Something clicked. I ordered a Mac Mini that week and started building the lab.
The first thing the lab taught me
I came in with ideas. More than a dozen specific projects I wanted to run — research questions, engineering problems, things I’d wanted to build for years. What became clear almost immediately was that without proper structure, a team of agents would create more work than it resolved. Each project had its own context, its own tools, its own outputs. Without a rigorous way to define scope, memory, authority, and handoffs, agents had no way to know what they were supposed to do, let alone coordinate across projects.
Harness engineering became the first serious work of the lab almost by necessity. As I built it out, I found the broader community was working toward the same problems from different directions — prompt engineering, context architecture, scaffolding, guardrails — all named separately, all pointing at the same underlying discipline. The first technical article on this site is that framework: an attempt to name the whole (at least for the lab), not just the parts.
Why build a website around it
Two reasons:
The first is accountability. I don’t owe anyone output here. No deadline, no manager, no deliverable I’m responsible for. That freedom is real, but it cuts both ways — without some external orientation, a project like this can stall indefinitely. Publishing forces a standard. It keeps the lab moving and keeps me honest about what I’m actually producing versus what I’m just thinking about.
The second is the narrative around AI — and why it became one of the questions I couldn’t leave alone. The fear-mongering and the rushed commercialization look like opposites but they’re two sides of the same problem: neither is actually engaging with what the technology is or how to use it responsibly. When you paint your house with your eyes closed, it’s not the paint’s fault the walls are a wreck. The right answer to that isn’t an argument. It’s a demonstration.
That’s what this site is. Not a claim that human-AI teaming works — a record of actually doing it, in public, with the reasoning visible and the failures included. What lives here is the work and curiosity that drove it. Hopefully readers will find it interesting too.
The questions that filled my head for years finally have somewhere to go. That’s what the lab is for.