Essay

Discipline belongs in the code, not your willpower

2026

I learned this the expensive way, so let me give you the number first: I lost about $21,000 because I believed I could control myself in the moment. I made it in three months and gave all of it back, bag-holding a position through a downturn that I was sure would reverse. It didn’t, not in time. The thing that broke wasn’t my analysis. It was the assumption that the disciplined version of me would be the one sitting at the desk when it mattered.

He was not. The person at the desk was tired, certain, and emotionally committed to being right. He had excellent reasons for every rule he broke. That’s the part most people get wrong about discipline failures — they imagine a moment of weakness, a lapse. It isn’t a lapse. It’s a calm, articulate override, delivered by a version of you who genuinely believes the rule doesn’t apply this time. You cannot out-willpower that person, because that person is your willpower, and it has been quietly recruited to the other side.

A while later I had the lesson hammered in again, harder. I’d built an automated system, but I’d left myself a manual override — a way to step in and “fix” things. A position ran against me, I intervened the way I always had, and a broker force-liquidated the account. I had automated everything except the one component that kept failing: me.

That’s when the principle finally landed, and it’s the principle I now build everything around: when a human is the unreliable part of a loop, you don’t repair the human with resolve. You move the constraint outside the human, where it can’t be reasoned away in the moment.

In my trading agents now, the discipline doesn’t live in my head. Position sizing is a Kelly-fraction calculation — a formula, not a conviction level. The risk rules are a written Investment Policy Statement, and that document is not something I reread for inspiration; it’s enforced in code. The piece I rely on most: when the system breaches its own limits, it automatically reverts to simulation before I’m given the chance to override it. The off switch does not ask my permission. That’s the entire point. The version of me who wants to flip it back on is precisely the failure mode the switch exists to stop, so handing him the key would defeat it.

People sometimes read this as a lack of self-trust, like I’ve given up on becoming disciplined. It’s the opposite. Building the constraint into the machine is what discipline actually is once you stop romanticizing it. A pilot doesn’t trust himself to remember the pre-flight checks; he runs the checklist. A surgical team doesn’t trust itself to count the instruments; the count is a procedure. Maturity isn’t trusting yourself more. It’s arranging the world so that less depends on you being at your best.

I think this generalizes well past trading, and it’s where my interest in autonomous systems and AI safety actually comes from — not from theory, but from having been force-liquidated by my own override. An AI agent that can place orders, run shell commands, or take real actions in the world has the same structural problem I did. You cannot make it safe by giving it better intentions or a more thoughtful prompt, any more than I could have willpowered my way out of that drawdown. It needs limits it cannot argue itself past — hard external boundaries that hold even when the system has constructed a perfectly reasonable-sounding case for crossing them. The good reasons are the dangerous part. A constraint that yields to a good-enough argument isn’t a constraint; it’s a suggestion.

And the human supervising that agent needs the same treatment, for the same reason. At 2am, tired and certain, the person watching the autonomous system is the same weak link I was at the trading desk. “A human will catch it” is willpower wearing a lab coat. If catching it matters, it has to be a mechanism, not a hope.

So the question I’d leave you with is the one I had to answer with real money: where does your discipline actually live? If the honest answer is “in my good intentions” or “someone will be paying attention” — that’s not discipline. That’s a story you tell yourself between the times it fails. Put it somewhere it can’t get tired. Put it in the code.