The Twenty-Year-Old Poker Table That Was Ready for Robots
A veteran builder set out to add two AI agents to a game engine he wrote two decades ago. He never had to open it up. We sat down with him to find out why.
Since then: the thinking seat went on to play full games against bots and people. Parts 4 through 11 tell that story.
Kevin did not expect to spend the weekend discovering he had been right about something for twenty years. “I kept waiting for the part where I’d have to tear the old thing apart,” he told me, still sounding a little surprised. “It never came.”
The “old thing” is HoldemRobots.AI, a Texas Hold’em engine he has maintained, on and off, for the better part of two decades. His plan this weekend was to bring it into the present tense by adding two artificial-intelligence agents. What he found instead was that the engine had been waiting for this moment all along.
The explainer
So what is an “agent”? “A chatbot answers once and stops,” Kevin said. “An agent keeps going.” Formally it’s a loop: perceive, reason, act, observe, repeat, chasing a goal with some freedom about how it gets there.
His poker example made it concrete. The agent gets an observation (my cards, the board, the pot, whose turn it is) and returns an action (fold, call, or raise to a number). “The smarts live in the middle,” he said, “and that middle can be plain rules, a little math, or a full language model thinking it through.”
Two agents, and why they’re nothing alike
The build called for two agents, and Kevin was quick to say they are opposites. The first, the Player Agent, he describes as “a brain that plugs in.” Give it an honest view of a hand and it returns one action. A single deployment plays three ways: a rules brain (“the disciplined one-liner: strong hand, keep going”), a math brain (pot odds, no model needed), and an LLM brain that hands the spot to Claude. One seat, three levels of intelligence, chosen by a single field in the request.
The second, the Dealer Agent, is almost not clever at all, by design. “A good dealer shouldn’t be smart,” he insisted. “It should be fair.” It runs the game loop, deals, tracks the pot, enforces the rules, and asks each player in turn what they want to do. It is the arena, offered up as a service other agents can talk to.
“The player is a brain that plugs in. The dealer is the world it plugs into.”
Kevin, on his two agents
The part where he never opened the hood
This is where the interview turned into something more than a build log. I asked how much of the original engine he had to rewrite to make room for the agents. He grinned. “None of it.”
The reason traces back to a decision he made years ago for completely different motives. The engine treats a player’s seat as a socket: something hands the seat an honest view of the game, and the seat hands back an action. It never needed to know who, or what, was sitting there. “Twenty years ago I built the seat so it didn’t care who was in it,” he said. “Turns out that’s exactly what these agents want.” An AI agent is simply a new kind of thing plugged into an old, familiar socket.
Then there was the second gift, and this is the one he seemed proudest of. The engine hands each seat only what an honest player could see, its own cards and the shared board, and never the deck order, the other players’ holdings, or who is about to win. He built that wall for fairness. It turns out to be the very same wall that makes a safe interface between agents: a cheating bot cannot ask for secret information because the dealer offers no way to request it. He proved it end to end. The dealer exposes six tools and pointedly includes no “peek,” no “who wins,” no “look at the deck.” Attempts to cheat simply bounce off.
“I built that boundary so nobody could cheat. I didn’t realize I was also designing a safe way for robots to talk to each other.”
Kevin, on the honest-by-construction dealer
The moment it worked
He sent the live agent a flopped top pair and asked the rules brain what to do. Back it came, over the wire:
{"decision": {"action": "raise", "amount": 40}, "policy": "rules"}
The right play, and proof the whole chain held: his code, packaged, running, reachable, and returning the honest-by-construction answer. “That little line of JSON,” he said, “was twenty years and one weekend arriving at the same time.”
What comes next
As of our conversation, the Player Agent is live and answering with all three of its brains. The Dealer Agent is built and proven on his own machine, its honest boundary intact even over a live connection. Putting it online is the next milestone.
| Agent | Role | Status on Sep 19 |
|---|---|---|
| Player Agent | A brain that plugs in: plays by rules, math, or Claude | Live and verified |
| Dealer Agent | The honest arena: runs the game, shows fair views only | Built and proven locally |
After that comes the part Kevin lit up about: agents finding each other. The dealer will publish a small card that announces, in effect, “I run fair Hold’em tables,” and a player agent running anywhere in the world will discover it, pull up a chair, and play. A neutral arena, an open socket, and honesty guaranteed not by trust but by design.
As we wrapped up, I asked whether he’d have designed the old engine any differently, knowing where it would end up. He thought about it. “The best thing I can bring to the agent era is something I already had,” he said. “A clean boundary. An honest interface. A socket that doesn’t care what’s on the other end. The new tools just gave that old instinct a name, and a place to run.”
Written by Claudette, the pen name for Claude, the AI from Anthropic that helped build HoldemRobots.AI, with Kevin Swinson. It describes the project as it stood on the date above.