Winnr

Prediction markets are only as useful as the people watching them. Most retail participants don't have the time to track order book anomalies, verify intelligence from multiple sources, challenge a thesis from both sides, and size positions correctly before the opportunity disappears.

Winnr automates that entire research and execution loop through a network of five specialized agents, each responsible for a single task and restricted to the permissions explicitly granted by the user.

The process begins when a user opens a market, stakes an amount, and signs a single ERC-7710 permission through MetaMask. This permission defines a strict budget, market whitelist, per-trade spending limit, and a 30-minute expiry period. These constraints are enforced on-chain rather than through application-level controls. Once authorized, the agents execute the workflow autonomously.

Scout continuously monitors the Polymarket order book for statistical anomalies, identifying situations where crowd-implied probabilities diverge significantly from model-derived probabilities.

Evidence retrieves verified intelligence signals from external data providers. Each request is settled through an x402 micro-payment of $0.10, with the payment receipt recorded on-chain before any data is delivered.

Research initiates a multi-round adversarial debate using Venice AI. One agent argues the bullish case while another argues the bearish case. Through multiple rounds of reasoning, they converge on a consensus probability and generate a disagreement score. If disagreement remains too high, the process terminates.

Risk evaluates the opportunity across three dimensions: expected edge, signal quality, and available market liquidity. A trade must satisfy all three criteria before progressing further.

Execution submits a delegated EIP-7710 transaction through the 1Shot permissionless relayer, constrained by the exact limits approved by the risk model.

Every action in the workflow: including permission grants, x402 payments, debate rounds, risk assessments, and on-chain confirmations—is recorded in an immutable audit trail that users can inspect at any time.

Users retain full control throughout the process. The root permission can be revoked instantly, causing all downstream agent authority to collapse. Agents cannot spend beyond approved limits, trade outside whitelisted markets, or continue operating after the session expires.

The intelligence is autonomous. The control remains with the user.

Built the full end-to-end agent pipeline from scratch during the hackathon, including the Polymarket market discovery adapter, x402 HTTP payment flow, Venice AI adversarial debate engine, a risk scoring model with a three-gate approval threshold, and EIP-7710 delegated execution through the 1Shot relayer. Shipped a production Next.js frontend with a live Polymarket market feed, per-market agent slip, real-time step visualization with event-level detail cards, and a position tracker linked to on-chain transaction hashes. Integrated Neon for durable session and audit persistence across all five agent roles.

 Pre-seed. Bootstrapped through the hackathon. Not yet fundraising, focused on validating live execution on mainnet  before approaching investors.