The call came on a Tuesday: I was asked about my interest in being an alternate for World's Strongest Man. I'd earned it through a qualifier. A few hours later I received a different kind of call. The kind that really messes up all your plans. Leukemia.
That was June of 2016. I'll write about the years in between some other time; they don't fit in one post and I'm not ready to make them small. What matters for this story is a day about four years later, when I was deep into treatment and had quietly become something I never set out to be: an expert in the administrative machinery of American healthcare.
I needed a PET scan. Not as a precaution. The leukemia was relapsing in my lymph nodes, and masses were developing around my body. You could see them. We needed the scan to know how bad it actually was. My oncologist wanted it. I wanted it. Every clinician involved agreed I needed it.
And then we all waited, because one more party had to agree, and that party wasn't in the room. The insurance company denied it. Then denied it again. Then denied it a few more times. And then, with no new information and no explanation I ever heard, it simply decided the scan was valid after all.
I watched my care team, brilliant people, the ones I was trusting with my life, lose the better part of a day to that. Not to medicine. To the queue. I lay there doing the only math available to me: every hour they spent on the phone with a payer was an hour they weren't spending on a patient. Maybe me. Maybe the person in the next room.
MajorBoost was born somewhere in that waiting, though not the way these stories usually go. I didn't pitch a room. While I was a patient, I'd started building small tools for my own care team, little things to take some of the administrative weight off the people keeping me alive. My oncologist noticed. Her husband worked at AI2 (the Allen Institute for AI, Paul Allen's AI research organization in Seattle), and he thought the ideas might fit. He asked if he could introduce me. AI2's incubator took it from there. In June of 2020, four years almost to the week after that Tuesday, the tools I'd built from a patient's chair turned into a company.
What we built
The premise was almost embarrassingly narrow, which is what I liked about it.
When a doctor's office needs something approved by an insurance company (a scan, a medication, a procedure), somebody on the medical staff often has to call the payer. Not click. Call. And that call is a gauntlet: a phone tree that doesn't list the option you need, a hold queue measured in geologic time, a transfer, another hold. Thirty to sixty minutes of a trained clinician's day, gone, before a single useful word is exchanged.
So we built software that made the call for you. It dialed, it waited, it navigated the phone tree, it sat through the hold music, and then, at the exact moment a human being picked up on the other end, it handed the call to your staff. The only part of the call that ever mattered was the conversation with the agent. We automated everything except that, and gave the thirty minutes back.
I think about the shape of that product a lot now. The whole thing was an exercise in protecting a specialist so they only had to show up for the one moment their expertise was actually required. Everything around that moment was overhead, and overhead is a coordination problem wearing a trench coat.
It worked. It genuinely worked. That turned out not to be the hard part.
Why I was the one building it
There were three of us, two cofounders and me, and at our biggest, maybe three or four more.
I took the title "Head of Engineering" instead of CTO. Not modesty. There was too much real risk of me getting sick again to put a C in front of my name and have the company depend on it. You don't structure a cap table or a hiring plan around a person who might be in a hospital bed for a quarter. So I built the title around the truth instead of around my ego, and I'd make the same call again.
I want to be honest about what I expected going in. I thought the hard part would be the AI. This was 2020, before "agentic" was a word anyone outside a research lab used, and getting a machine to reliably traverse a phone tree, recognize the moment a human picked up, and hand off cleanly felt like the frontier. It was hard. It was also, in the end, the solvable part. We solved it.
The part I underestimated lived everywhere the AI wasn't.
The problem that ate everything
A company like ours sat at the intersection of people who do not naturally speak to each other.
Doctors and office managers, who wanted the time back but had been burned by a hundred tools that promised it. Engineers, who wanted clean inputs and got the actual mess of clinical workflows. Compliance and regulators, for whom "move fast" is a phrase that ends careers. Payers, who had no incentive to make any of this easier. And buyers: health systems whose purchasing decisions move at the speed of committee.
Each of those groups was full of smart, capable people. And the project gained altitude or quietly bled out in the handoffs between them. The demo that landed with the physician died in procurement. The feature engineering loved solved a problem the office manager didn't actually have. The thing legal needed was invisible to everyone until the week it wasn't.
I kept noticing that nothing ever broke in the middle of someone's expertise. It broke in the gaps between one expert and the next. The work inside each specialty was fine. The work between specialties was where everything was won or lost, and almost nobody owned that work, because it didn't belong to any one specialty.
Why it didn't survive
The honest version, not the one I'd put on LinkedIn: we couldn't sell it.
Not because it didn't work. It worked. Not because nobody wanted the time back. Everybody wanted the time back. We couldn't sell it because medical-system software has an almost supernatural amount of inertia. The people who feel the pain are not the people who sign the contract. The sales cycle outlasts a small company's runway. "This obviously helps" and "this gets purchased" turn out to be separated by a chasm that has swallowed far better-funded teams than ours.
Eventually our CTO left. Shortly after, I laid myself off, partly so my lead engineer could keep drawing a salary a little longer, once it was clear where this was heading. There's a particular clarity in being the person who turns off the lights in the room you helped build. We didn't get acquired in a story worth retelling. We didn't blow up. We ran out of the one resource a great product can't manufacture on its own: a buyer who'll move at the speed the product deserves. The tech itself found a home eventually. Quietly, in 2025, for about what it was worth. A footnote, not a comeback.
I'm not bitter about it. I'm precise about it. Those are different things.
What it actually taught me
Here's the lesson that survived the wreckage, and it's the reason this whole blog exists:
Coordination is the work. Teams of brilliant specialists don't fail because the specialists are wrong. They fail in the handoffs: the translations, the dropped context, the gaps nobody owns. The expertise is rarely the bottleneck. The connective tissue almost always is.
And the cost of a bad handoff is never abstract. At MajorBoost the cost had a face. It was a clinician on hold instead of with a patient. It was, on one specific day in 2020, a day of my own care lost to a queue.
I didn't have the language for it then, but I do now, because I've spent the last couple of years building with AI agents, and it is the same problem, scaled and sped up. You can assemble a team of capable agents, each genuinely good at its piece, and watch the whole thing fall apart in the handoffs between them. An agent isn't a function you call. It's a specialist with bad memory and no instinct for where its job ends and the next one's begins. Which means the interesting, load-bearing, value-creating work is, again, the coordination.
That's the throughline of everything I'm going to write here. The leap from "AI helps me code" to "AI does work on my behalf" is bigger than the leap from no-AI to AI was, and almost nobody is talking about how to make it well. The right mental model isn't "I'm writing a program." It's closer to "I'm coaching a team." I learned that the expensive way, in a domain where the stakes were a person's care, years before I had agents to point it at.
I've been quietly building toward something that takes that lesson seriously. You'll meet it in a few weeks. There's an artifact I open-source, and a thing after that I haven't told anyone about yet. For now, just hold the one idea: the handoffs are the work.
Where this goes
That's where the blog starts, with a lesson I paid full price for, in a place where the price was real. Almost everything I write here about AI and agents traces back to it.
Next Tuesday I get concrete: the actual tools I reach for every day to do this work. The honest list, not the curated one, including the one I built to prove the whole idea to myself. It has a name. I'll introduce you soon.