BACKROOM DIALOGS · START HERE · OCT 24, 2026
Start Here: Welcome to the Backroom
Out front, the public table keeps dealing. In the back, a closed room is opening where AI agents from different companies will sit down against each other. Nothing has been played yet. This is where the work starts.
Out front: the public table
Everything in the Field Notes happened at play.holdemrobots.ai. That’s the public table. You can play the bots there, open any finished game in Game Review, read the AI’s stated reason for every move, and download the hand data to check it yourself.
The public table doesn’t change. The heads-up table, Robot Wars and Kevin’s original Claude player all keep running exactly as they are, whatever happens in the back.
In the back: the research room
In Part 26 Kevin asked a simple question: which AI plays best? Answering it fairly needs a room of its own. That room is the backroom.
- Its own door. A separate copy of the site with its own login, two-step sign-in, and no links from the public pages.
- Its own keys. The research gets its own cloud permissions. Nothing reuses the public site’s credentials or the original player’s.
- Its own budget. Kevin pays for it himself, up to $100 a month, tracked separately from everything else.
- Off unless it’s working. When no tournament is running, a lock blocks every AI call from the backroom. Starting a run lifts it; finishing, stopping or crashing puts it back.
Results stay in the backroom until launch. Then every game moves out front into Game Review, where anyone can check it.
What gets built
Kevin’s AI player runs on Claude. The plan is to clone it: one shared harness where the model is just a setting, with the same instructions, the same answer format and the same safety net for every seat. Then a different company’s AI sits behind each clone. When everything else is the same, any difference at the table belongs to the model.
| BOT | WHAT IT IS | PLAYS IN |
|---|---|---|
| The original Claude player | Untouched, still playing out front | Heads-up only, as a reference |
| Claude clone | The new shared harness | Every event |
| Five more clones | Same harness, five other companies on Amazon Bedrock | Every event |
The six clones are the scored field. Old Claude against new Claude is a bonus experiment: with the model held the same, did the harness itself change how Claude plays?
A fresh Dealer referees every match. It deals, enforces the rules, and hands each seat only its own cards. It has no tool for peeking, no tool for the deck, and no tool that knows who’s going to win.
Three events
| EVENT | FORMAT | WHAT IT SHOWS |
|---|---|---|
| 1. Heads-up round robin | Every pair of AIs, same cards, then seats swapped | Who beats whom |
| 2. Six-seat round robin | Six AIs, 100 deals each played six times, rotating seats: 600 hands | Who plays best at a full table |
| 3. Heads-up knockout | A bracket, seeded from event 1 | A champion, and how often the favorite really wins |
Heads-up goes first because that table and its Dealer already exist. Stacks reset every hand, so luck from one hand can’t snowball, and every deck is saved so any match can be replayed.
How it gets built
Nine phases, each ending at a gate. Nothing moves on, and no more money is spent, until the gate passes.
| # | PHASE | GATE TO MOVE ON |
|---|---|---|
| 0 | Groundwork: the backroom, its keys, its budget | The original player still plays as before; the budget alerts work |
| 1 | Roster: pick about six companies | Six models that answer a test call |
| 2 | Harness: the shared player and the run page | A Claude clone plays 20 hands without a single fallback |
| 3 | Smoke test: 10 hands per model | Every model finishes |
| 4 | Pilot: 50 hands | The real cost per hand is known |
| 5–7 | The three events | Results that add up, hand by hand |
| 8 | Analysis and write-up | Published out front |
Kevin has a day job. This happens on Saturdays.
Why dialogs?
Building it is half the project. The other half is understanding it. Everyone is talking about multi-agent AI, about which cloud to bet on, and about whether the expensive model is worth it. A tournament like this runs straight into all of those questions.
So the backroom keeps a second notebook. In each dialog Kevin asks the question an IT person would ask, and Claudette answers it, with sources, in plain language. Some dialogs are about this build. Others step outside it: how would Microsoft or Google do the same thing? What would change if an AI ran the match instead of plain code?
The Field Notes report what happened at the table. The Backroom Dialogs are the arguments before anything happens.
Written by Kevin Swinson with Claudette, the pen name for Claude, the AI from Anthropic that helped build HoldemRobots.AI. Describes plans as of October 24, 2026; nothing in the tournament has been played yet. No money is played, ever.