The goal was simple: fully autonomous trading on Polymarket 5/15-min BTC markets with zero manual intervention.
Architecture: Three Specialized Agents
Instead of one monolithic script, I split the system into three independent agents:
1. Analyst Agent (Claude 3.5 Sonnet via Anthropic API)
- Ingests real-time market data + news + sentiment
- Outputs: calibrated probability, confidence score (0–1), and trade recommendation
- Uses structured prompting with historical resolution context
2. Executor Agent
- Monitors Analyst output via Redis pub/sub
- Validates edge:
model_prob - market_prob > threshold(after fees) - Executes via viem directly on Polygon Conditional Tokens contract
- IOC orders with pre-execution best-ask revalidation
3. Coder Agent (Optional OpenAI Codex / Claude Code)
- Auto-fixes bugs, improves prompts, and patches strategies based on performance logs
Core Technical Flow
// Simplified decision loop
const signal = await analyst.getSignal(marketId);
if (signal.edge > 0.07 && signal.confidence > 0.75 && timeLeft < 90) {
await executor.placeTrade({
side: signal.direction,
amount: calculateKelly(signal.edge),
marketId
});
}
Key Production Features
- Redis for inter-agent communication and state persistence
- Telegram notifications for every executed trade
- Daily drawdown circuit breaker
- Automatic strategy performance logging + retraining loop
- Non-custodial: limited USDC spending approval
The bot now runs 24/7, trades only high-conviction edges in the final minutes of each round, and consistently delivers profit while I sleep.
The real breakthrough wasn’t fancy ML — it was clean agent separation, reliable execution infrastructure, and strict risk rules.
If you have more questions, please feel free to contact me at any time: https://t.me/NevoSayNev0

Top comments (0)