Descriptions:
Inspired by Andrej Karpathy’s autoresearch project, the creator behind the All About AI channel adapts the concept into a self-improving trading bot for Polymarket’s Bitcoin up/down prediction markets. Rather than training a neural network, this implementation uses Claude Code as an autonomous agent that iterates on trading strategy code stored in a GitHub repository, guided by a markdown file called the training program that defines how experiments are chosen, executed, and evaluated.
The system runs on one-hour experiment windows. After each window, a scoring function evaluates whether the new strategy improved on the current best—measured by metrics like fill rate and win rate on Bitcoin arbitrage trades (simultaneously buying up and down positions at a combined cost below 100 cents to capture spread). Strong results trigger a confirmation run before being committed; weak results are discarded. GitHub serves as both the version control layer and the agent’s cumulative research memory.
The video walks through the custom dashboard, live experiment transitions, Claude Code’s autonomous strategy update commits, and a final live test against real Polymarket positions. The architecture is a practical demonstration of applying self-improvement loops—normally seen in ML research—to financial prediction markets, with Claude Code acting as the autonomous research and engineering engine. Developers interested in agentic systems, trading bots, or self-modifying code loops will find the design patterns here directly applicable.
📺 Source: All About AI · Published March 11, 2026
🏷️ Format: Hands On Build
