George has played 1.5 million poker hands. Now he's building a company where luck is the enemy.
George Kurdin joined Not Another Podcast to talk about the road from professional poker to accounts receivable. Here is what stayed with us, and why it maps to how we build Monk.
Before Monk, George played more than 1.5 million hands of professional poker. His longest session ran about 60 hours. Poker is a game of luck, and six years at the table taught him to live with it.
Then he built a company in accounts receivable, where there is almost no room for luck. A single wrong move touches someone's real money and a real customer relationship. Poker taught him to make peace with variance. Accounts receivable does not allow for it.
Why accounts receivable
AR is not the flashy end of software, and George is the first to say so: "It's not that sexy, but it's very important for the world." What drew him was the standard the work demands. When you handle money for a business, the bar for getting it right is absolute, and clearing it becomes a moat. As he put it, "you earn the right to build a big business if you touch money."
What a million hands teaches
Most people assume poker is about reading faces. George points somewhere else. He says the part that carries over is "your relationship with risk, your relationship with expected value, discipline." And speed. With so many hands and immediate consequences, "you have to iterate quite quickly, otherwise you get crushed." That habit of iterating fast is the same one that makes software reliable.
The agent harness
The part that maps most directly to how we build came when the host asked George to define an "agent harness" in plain terms. His answer: a set of guardrails "the team builds around the model call" so the outcome is "resilient and not brittle."
In practice that means the model does not get the first move. Before Julia, our collections agent, writes a single email, deterministic workflows run first. The model works inside a space we have already decided is safe.
George reached for self-driving to explain it. Monk today is like an L4 car. The model does the driving, a human is there "to correct sometimes," and deterministic guardrails act like the lane lines and lights that keep the car on the road. Knowing what to hand to that human takes judgment. In his words, "it's an art to know when to escalate."
The approach he warns against is the black box that promises "we do it all for you" with "nothing to trace." For work that touches a customer's money, that "is not resilient enough." Every move Julia makes is logged, and a person keeps the final say. We wrote more about that standard in Boring agents by design.
There is a business reason to build this way, not just a safety one. George's bet is that the teams who get the harness right win on margin, because a reliable agent needs "a lower ratio of people to an outcome."
The moat when everyone calls the same model
If every company can call the same API, what is left to defend? George's answer was short: "brand and distribution," plus the last mile of a real workflow, the part that is hard to get right and where he believes the value sits.
That is also why the team stays small and selective. Fewer than 1 in 700 applicants are hired. Monk runs on one meeting a week, no standups, and whoever owns a piece of the product gets to make the call and ship it. Reliability is easier to hold when a small group of people own their work end to end.
Watch the episode
George covered far more with Not Another Podcast, from the labs to the future of white-collar work. The full conversation is worth your time.
.avif)