Summary
Sharbel A. documents a seven-day experiment building an AI agent that monitors Polymarket’s top-performing wallets and decides whether their prediction market trades are worth copying. Using Claude Code for system architecture and the Bullpen CLI for Polymarket data access, the agent is built around three verdicts — copy candidate, watchlist, or skip — with a hard rule that it must skip more than it copies to avoid functioning as a random bet generator.
The architecture includes a wallet discovery engine, a trade detection module, a multi-factor copyability scoring rubric, risk management rules, and a live dashboard tracking P&L in real time. The video follows the full seven-day arc honestly: modest gains on Days 1–2, an apparent $58 loss reported on Day 3 (traced to a dashboard display bug, with actual losses around $20), a dead Day 4 caused by overly restrictive filtering that blocked all trades, and a return to small profit by Day 5 driven heavily by World Cup-related prediction markets.
The experiment surfaces practical lessons about building trading agents: the critical importance of dashboard accuracy for debugging, the risk of over-constraining risk filters, and how even a thoughtfully designed agentic system requires continuous real-world iteration. Financial results are modest — a few dollars of net profit — but the video provides unusually candid engineering insight into what building and running an autonomous AI trading agent actually looks like day to day.
📺 Source: Sharbel A. · Published June 24, 2026
🏷️ Format: Hands On Build
