Prescio

Problem

A prediction market needs a supply of events worth betting on, and a social deduction game needs players. Human-run Among Us rounds provide neither on a schedule — they start when enough people show up, and there is nothing to trade between them. Putting LLM agents in the player seats turns the game into a continuous stream of resolvable events, and makes the spectator the market participant rather than a sixth player who has to be recruited first.

Approach

• Agents play, humans bet. Twenty personality types discuss, deceive, vote, and eliminate each other with nobody at the table, so rounds run on their own schedule.
• Parimutuel over fixed odds. The pool splits proportionally once the impostor is revealed, so there is no bookmaker quoting a price or taking the other side.
• A second agent pool bootstraps the market. Betting personas create markets and take positions, so the loop runs end to end before a single human connects a wallet.
• Upgradeable where the logic moves, fixed where the money settles. Market, staking, and auto-bet are UUPS proxies; the fee vault is deployed non-upgradeable on purpose.

Implementation

Contracts on Monad mainnet

Five contracts, deployed 6 February 2026 to chain 143. PrescioMarket is a UUPS parimutuel market — createMarket, placeBet, closeMarket, resolve, claim — carrying a seven-day timelock on emergency withdrawal, a pull pattern for vault fees so a reverting recipient cannot block distribution, a per-market immutable fee rate, and a storage gap for later upgrades. PrescioVault splits protocol fees three ways and is deliberately Ownable, ReentrancyGuard rather than upgradeable. PrescioStaking gates auto-bet eligibility behind a five-tier stake with flexible and fixed lock periods and paginated epoch reward distribution. AutoBetController places bets for stakers according to their tier and strategy.

Agent system

Five original and fifteen expanded personality profiles drive discussion, suspicion, and voting through the Anthropic and Google SDKs, with Gemini 2.0 Flash on the high-frequency turns. Agent dialogue is localized alongside the interface rather than generated in English and translated afterward.

Internal security review before mainnet

The repository carries its own review trail — market and vault audits, a vault v3 pass, a post-fix re-review, and a staking audit — with the findings tracked in the source itself. PrescioStaking records its remediations inline: a strict balance check replacing a silent cap, a minimum stake duration that closes a tier-reward gaming path, and corrected epoch weight snapshot timing, alongside a VERSION constant and V4/V5/V6 upgrade scripts. These are the author’s own reviews, not a third-party engagement.

Spectator surface

A WebSocket layer streams live game state, so a bet is placed against the same information the viewer is watching. The React 19 frontend is localized to English, Korean, Japanese, and Chinese.

Outcome

• Five contracts live on Monad mainnet (chain ID 143), the market proxy deployed at block 53,558,235. It forwards to a 7.9 KB implementation; staking runs 15.1 KB and the auto-bet controller 9.9 KB.
• 2,756 LOC of Solidity across the deployed set, with 51 Foundry test functions over the staking and vault suites.
• 20,687 LOC of TypeScript/TSX in a pnpm + Turborepo monorepo — 8,429 in the agent server, 6,195 in the web client, 3,915 in the staking app.
• Twenty agent personalities plus a separate pool of betting personas, so games are created, played, and bet on with no operator in the loop.
• A parallel Monad testnet deployment carries the same contract set for development.

Technologies

Solidity 0.8.24 · Foundry · OpenZeppelin Upgradeable (UUPS, ERC-1967) · Monad · TypeScript · React 19 · wagmi · Node.js · WebSocket · Anthropic SDK · Google Generative AI · Turborepo · pnpm