Edge Scout

Inspiration

Prediction-market traders drown in noise: thousands of markets, wallets moving constantly, and no honest signal for which move is actually smart money versus a whale making a bad bet. Watching flow all day is a full-time job a small trader can't afford to hire for. Edge Scout is that analyst.

What it does

Edge Scout runs quietly in the background and pings the trader only when there is a real, track-record-backed edge the price has not already caught. Most scans end in silence, which is the point.

• Anchors on Predge's signed smart-money consensus: net USDC flow and direction from wallets scoring above 70 on realized win rate.
• Confirms candidates with whale concentration and per-wallet track records, and rejects moves that already ran.
• Emits a single conviction-scored alert, or holds. Alerts go to a webhook (Telegram, Slack, or a backend) so it truly runs in the background.

How we built it

• Strands Agents SDK for the agent loop and six @tool functions.
• Predge (api.predge.io) as the data layer: a production x402 pay-per-call API over Polymarket smart-money data.
• x402 + eth-account for payment: each tool call hits a 402, signs a USDC micropayment on Base, and retries. Predge settles only on a 2xx, so a failed call costs nothing.
• Model: Amazon Bedrock Claude by default (Anthropic and OpenAI are drop-in). Ships with a one-file Bedrock AgentCore entrypoint.
• A MOCK mode (sample data, no wallet) so anyone can clone and run it instantly.

The agent-economy angle

This is not just "an LLM calls an API". The agent autonomously decides which one question it needs answered and buys exactly that data for a few cents: a real agent-to-service micro-transaction happening inside a single tool call.

Challenges we ran into

Tuning the agent to stay quiet. The first version alerted on everything. The fix was hard conviction gates in the system prompt (one-sided flow AND proven wallets AND unspent price) plus a numeric threshold in the watcher, and keeping the repo runnable with zero setup while still demonstrating real payments (the MOCK/LIVE split).

What's next

Position sizing from the trader's own risk limits; always-on hosting on Bedrock AgentCore Runtime; and Predge signed-call attestations so every alert is verifiable after the outcome resolves.

Built With

• ai-agents
• amazon-bedrock
• base
• claude
• eth-account
• polymarket
• predge
• python
• strands-agents
• usdc
• x402