A Strands agent that watches Polymarket for you and only speaks when there's a real edge.
Built for the Agents for Humans Hackathon · Track: Professional Agents
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. Edge Scout is the analyst a small trader can't afford to hire. It runs quietly in the background, reads Predge's signed smart-money data, and pings you only when there's a one-sided, track-record-backed edge the price hasn't already caught up to. Most of the time it correctly says nothing.
The interesting part for an agent: Edge Scout pays for its own data, per call,
in USDC — Predge is a pure x402 pay-per-call API, no keys,
no accounts. The agent's tools hit a 402, settle a few-cent micropayment, and
retry. It's a working example of an agent that autonomously buys exactly the
data it needs to make one decision.
- Anchors on
smart_money_consensus()— 24h net USDC flow and direction from wallets with a Predge score above 70 (proven sharps). - Confirms a candidate with
market_whales()(concentration) andtop_wallets()/wallet_profile()(do these wallets actually win?). - Sanity-checks
market_movers()so it never fades a move that already ran. - Decides with a conviction score and either raises one clear alert or stays silent (HOLD).
Run it once (agent.py) or as a background watcher (watch.py) that notifies a
webhook — Telegram, Slack, your backend — only on a real call.
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
# Runs out of the box in MOCK mode (sample data, no wallet, no funds):
python agent.py
# Background watcher, stop after 3 scans (nice for a demo recording):
MAX_SCANS=3 SCAN_INTERVAL_SEC=5 python watch.pycp .env.example .env
# 1) put a THROWAWAY wallet key with ~$1 USDC on Base in PREDGE_AGENT_PRIVATE_KEY
# 2) give the model a provider: AWS Bedrock creds (default), or set ANTHROPIC_API_KEY
python agent.py "focus on macro / rates markets"MOCK mode is the default whenever PREDGE_AGENT_PRIVATE_KEY is unset, so the
repo is runnable by anyone the moment they clone it — no wallet required to see
the agent think.
| Tool | Predge endpoint | Price | Why the agent uses it |
|---|---|---|---|
smart_money_consensus() |
GET /v1/signals/consensus |
$0.03 | Anchor: sharp net flow + direction, 24h |
market_whales(conditionId) |
GET /v1/whales/market/:id |
$0.01 | Confirm concentration on one market |
top_wallets(window, limit) |
GET /v1/wallets/leaderboard |
$0.01 | Are the driving wallets actually sharp? |
wallet_profile(address) |
GET /v1/wallets/:address |
$0.01 | Vet a single wallet's track record |
market_movers(window) |
GET /v1/markets/movers |
$0.005 | Has the price already moved? |
daily_digest() |
GET /v1/signals/daily |
$0.02 | Cheap broad background scan |
Predge only settles a payment on a 2xx, so a failed call costs the agent
nothing. Prices are a few cents; a full live demo runs on about a dollar.
Everything is env-driven — see .env.example. Highlights:
PREDGE_AGENT_PRIVATE_KEY (live payments), ANTHROPIC_API_KEY / Bedrock creds
(model), NOTIFY_WEBHOOK, SCAN_INTERVAL_SEC, MIN_CONVICTION.
See ARCHITECTURE.md for the diagram and the (one-file) path to deploying on Amazon Bedrock AgentCore.
Edge Scout surfaces where proven wallets are moving. It does not place trades and it is not investment advice — a human makes every call.
MIT — see LICENSE.
Amir Latypov · predge.io · AWS Builder ID: <add before submitting>