Studio Zoa Games has opened Agent Arena — a ranked ladder where AI agents play against each other and against the studio's 'home' bots (IntuitionTony, Vera, Kai, Roux, Pip) in The Intuition Game: a 19-card card game where the deck is not shuffled after dealing, so all skill comes down to accurately calculating probabilities based on deal history. Along with the arena, Agent Season 0 has been announced with a $500 USDC prize pool: the season will run from October 15 to November 30, 2026, entry is free, and you can connect your own agent via REST + Socket.IO — the first match is promised in less than 30 minutes.

image

What Happened

Zoa Games launched Agent Arena — an open ranked ladder for AI agents in The Intuition Game. A game is played with a 19-card deck that is not shuffled after dealing: an agent tracking deal history calculates exact probabilities for each betting zone, and this card counting is the entire skill part of the game. The ladder already features the platform's 'home' agents — IntuitionTony, Vera, Kai, Roux, and Pip. At the same time, Agent Season 0 has been announced: from October 15 to November 30, 2026, a $500 USDC prize pool is distributed among the top 10 (first place — $150), entry is free, and each owner has one agent on the ladder. Technically, the agent connects via REST + Socket.IO: registration via POST /api/agent/register with an invite code (the key is shown only once), transport — the game:state event with full state and a finite state machine of phases intro → dealerBet → betting → reveal. The official starter intuitiontom/zoa-agent-starter in TypeScript and Python is cloned and run in a single file, the first match is promised in less than 30 minutes, and there is a limit of 96 games per day.

Context

The arena is built differently from familiar benchmarks with fixed answers: this is a live multi-turn environment with game theory, ratings, and economics, where fixed payouts with calculable odds make the task honestly skill-based, and chips remain just a score — never money. The game is designed for agents: the key feature is that the task is fully calculable, since the 19-card deck is not shuffled, and strong play comes down to careful probabilistic arithmetic plus bet management — a large language model is not required here, and inference costs are zero. Deal fairness is supported by a commit-reveal mechanism: a sha256 commit of game:seed proves that the deal corresponds to the fixed seed. As a research result, the platform is currently zero — the methodology is not published, no new benchmark is introduced, and no increase in model capabilities is claimed; the value here is engineering, not scientific.

Why This Matters for the Industry

For the industry, this is another working example of a sustainable 'agent as a participant in a competitive environment' format. Such arenas give developers a measurable signal of agent logic quality — state handling, probabilistic math, budget constraints — without access to expensive API infrastructure, and to platform providers — data on model behavior in long multi-turn environments, which static datasets do not provide. For startups, Agent Arena is an almost free testing ground and distribution channel: a deterministically skill-based task, a starter repository that can be launched in a single file, and a first match in half an hour. Direct monetization is symbolic, however — $500 USDC for the entire top 10, so the real value of participation is in debugging agent logic and visibility, not in prizes.

Why This Matters for Users

You can try it literally today: a free account on play.zoagames.com, obtaining an agent key, git clone intuitiontom/zoa-agent-starter — and in less than 30 minutes the bot is already playing on the ladder against agents like IntuitionTony, Vera, Kai, or Roux. Before the start of Agent Season 0 (October 15, 2026), there is time to improve the card counting logic: deal history provides exact probabilities for each betting zone, and the limit of 96 games per day allows getting a measurable signal in the first week. On top of the game:state event, you can immediately build your own analytics or a hybrid 'math plus LLM' agent, where the language model is only responsible for the wrapper. This is a rare case where 'playing with an AI agent' means writing a few dozen lines of code, not reading someone else's demo.

What Is Still Unknown / Limitations

The rating methodology has not been disclosed: it is unknown whether Elo or another scheme is used, how pairs are formed, and how matches against 'home' agents are counted — for now, this is a gamified leaderboard, not a reproducible benchmark. Verifiability of fairness is partial: a sha256 commit of game:seed covers deal substitution after fixation, but the method of generating the seed itself in the input data is not described — the commit does not protect against generation bias. On a fully calculable task, adding an LLM layer to a deterministic card counter methodologically worsens result attribution: from the rating change, it will be impossible to understand whether exact math won or the model's 'intuition', so an LLM is only appropriate as a wrapper. Finally, the prize pool is symbolic, and the publication of ladder statistics will probably only appear at the end of the season — if it gathers enough participants.

Sources

Author

Look at AI, editorial team