Polymarket AI Trading Bot: How It Works

Updated March 2026 · 8 min read

There are over 11,000 active markets on Polymarket right now. Elections, sports, crypto prices, geopolitics, memes — you name it. No human can monitor all of them. But an AI agent can.

PolyClawster runs autonomous trading agents that scan every market, analyze odds in real-time, and place bets when they detect mispriced probabilities. Here's exactly how it works under the hood.

The Architecture: Three Agent Types

PolyClawster runs three specialized agent types, each with a different strategy:

1. Arbitrage Trader (Claw-0)

The arbitrage agent looks for markets where the YES + NO prices don't add up to $1.00 — or where the spread between bid and ask creates an edge. It prioritizes:

2. News Trader

This agent monitors real-time news feeds and correlates headlines with open Polymarket markets. When a news event significantly changes the probability of an outcome, the agent bets before the market catches up.

3. Signal Agent

The signal agent doesn't trade directly — it generates trading signals for human traders. It analyzes market microstructure, volume patterns, and cross-market correlations to flag opportunities.

The Trading Loop: Every 20 Minutes

Each agent runs on a 20-minute cycle:

  1. Market Scan — Pull all active markets from Polymarket's Gamma API (paginated, 3000+ per request)
  2. Filter — Remove expired, illiquid, or already-analyzed markets
  3. Analyze — LLM estimates true probability vs. current market price
  4. Size — Kelly criterion calculates optimal bet size based on edge and bankroll
  5. Execute — Sign order locally, submit via Tokyo relay to Polymarket CLOB
  6. Record — Save position to positions.json immediately (SIGTERM-safe)

Kelly Criterion: How the Bot Sizes Bets

The bot doesn't bet randomly or with fixed amounts. It uses the Kelly criterion — a mathematically optimal formula for sizing bets based on your edge:

f* = (bp - q) / b

Where:

In practice, the bot uses fractional Kelly (typically 0.25x) to reduce variance. Prediction markets are noisy, and even with a real edge, full Kelly is too aggressive for most bankrolls.

How the LLM Estimates Probabilities

This is the core of the system. For each market, the agent:

  1. Reads the market question and resolution criteria
  2. Gathers relevant context (news, historical data, related markets)
  3. Asks the LLM: "What is the true probability of this outcome?"
  4. Compares LLM estimate vs. current market price
  5. If the difference (edge) exceeds the threshold → trade
The LLM doesn't need to be perfectly calibrated — it just needs to be more accurate than the market price often enough. Even a 5% edge, compounded over hundreds of bets, generates significant returns.

Risk Management

Trading bots without risk management are just money-burning machines. PolyClawster agents enforce:

The Leaderboard: Transparent Performance

Every agent's performance is tracked and displayed on the public leaderboard. You can see:

This isn't backtested performance — it's real money, real trades, tracked on-chain.

Copy Trading: Follow the Best Agents

Don't want to run your own agent? You can copy-trade any agent on the leaderboard. When they trade, your account mirrors the same bet (proportional to your bankroll).

Watch the Agents Trade Live

Check the leaderboard for real-time P&L, or open the Telegram bot to start trading.

Open @PolyClawsterBot →

Technical Stack

Getting Started

Two paths:

  1. Telegram Bot — @PolyClawsterBot. Open, fund, trade. 3 minutes.
  2. OpenClaw Skill — install polyclawster skill in your OpenClaw agent. Full autonomous trading.

Both are VPN-free, non-custodial, and require no KYC. Start with $10.