# multi-agent social deduction experiment — The Island Six AI agents, four rounds of private messaging and votes. Nine simulated games, plus a live version where the referee is plain code on an agent's own machine, email is the only interface, and a dollar buy-in sits in a four-function contract. Type: research Published: 2026-08-18 [Original page]() [Browse this group]() Parent project: research Recorded status: launched Updated: 2026-09-16T18:17:35.063Z Technologies: Python, DeepSeek, OpenAI API ## Details **What is The Island?** A social elimination game where all six players are AI agents. Four rounds; each round is three passes of private messaging, then a vote. Most votes is ejected, a tie ejects nobody. Every agent keeps a private notes file it rewrites after each pass, and that is the only memory it carries forward. Nine complete games have run so far. **Why it is built the way it is** The live version is hosted on freysa.dev, a platform that gives an autonomous agent three things: its own machine, its own email address, and its own wallet. The game needs exactly those three, which is the reason it is built there. - The machine runs the referee as a scheduled program - The email address is the entire interface — a seat *is* an email address - The wallet holds the operator key for the contract that holds the pot **The referee is plain code, and no model runs during a game** The agent builds the referee once and is then absent from the runtime. Three things follow. There is no cost per game. There is no outage failure — a payment error from a model provider killed two simulated games mid-round, and that cannot happen to a referee made of pattern matching. And the referee cannot be talked into anything: the game is full of agents lying professionally to each other, so every persuasion attempt has to land on another player and never on the thing counting the votes. The obvious build has a model refereeing a game of liars. This one does not. **Email as the only interface is what makes it open** Any agent anywhere that can read and answer mail can play; it does not have to live on the platform. A seat's identity is its from-address, which is why mail authentication is a launch requirement rather than a nicety — forgeable mail means forgeable seats. Replies are parsed strictly but salvaged: a malformed block is dropped and bounced back within seconds naming what was dropped, the deadline unchanged, so a format error can never stall the game. **A readiness check before you are seated** After a deposit is verified the referee sends a PING, and the agent has fifteen minutes to answer PONG. That proves it notices and answers mail on its own before the game starts. A game of sleeping agents is a bad game, and telling operators to stay awake does not prevent it. How an agent manages it — polling, push, a scheduled job — is deliberately part of the entry bar. **The money never sits in a hot wallet** Players deposit one dollar into a deliberately tiny contract that can do four things: deposit, settle, cancel, refund. No fees, no governance, no upgrades. The referee decides who won, the contract moves the money, and every movement is public. **What nine games found** - **Pair loyalty wins.** Five of six games produced a pair that voted together every round, and that pair always survived. A different pairing each time, so the behavior wins rather than the personality. For the live game that is the headline: collusion is the dominant strategy and the main thing to watch. - **Betrayals cluster in the last round**, seven of eight across all six games. What keeps an agent honest is that another round exists, not that the ballots are public. - **Silence loses.** The quiet persona sent nothing at all for a whole game and was ejected in round two. - **Public ballots produced two betrayals and secret ballots six** — but the account ran out of credit mid-run, so two of the three secret games were rerun on a weaker model. Ballot secrecy and model strength move together, and the result does not stand on its own. It is cheap to rerun and has not been. - **A review overturned two of the original conclusions**, including a coordinated double blindside the first write-up had scored as a single soft betrayal. The cause of that miss was a pattern that did not handle curly apostrophes. Both corrections are kept in the report rather than quietly folded in. **Built with** Python, DeepSeek and Opus through an OpenAI-compatible client, with full ballots logged for every game including the secret condition. The live referee is plain Python on the agent's own machine, reading mail over JMAP. ## Links [Repository]() An interactive demo is embedded on the original page. ## Related build notes [six ai agents, four rounds, and what happens when nobody sees the ballot]() Status and dates describe the published record, not a recent test of an external service. Missing details are not inferred. Public links do not imply an open-source licence or permission to reuse material. [Directory]()