Skip to content
Agentic MEV Day 2026

The Trading Agent Loop: How to build Claude Code for trading

Michael Feng

Session summary

The presenter describes Hummingbot as widely used open-source software for crypto algorithmic traders and market makers, noting that reported user activity includes tens of billions in trading volume across many exchanges and blockchains. He emphasizes that much of the connector and strategy ecosystem is built by the open-source community, and explains that the project’s funding comes from trading-fee donations from exchanges where users deploy Hummingbot. He also notes that AI tools are starting to bypass traditional software, affecting Hummingbot’s revenue, and frames this as a broader challenge facing open-source projects.

Drawing inspiration from what he calls the coding agent loop behind tools like Claude Code, the presenter proposes an analogous four-step “trading agent loop”: define a precise, testable strategy with parameters and risk limits; collect and transform data into signals; backtest on in-sample and out-of-sample sets; then transition cautiously into live trading with small size, monitoring and iterating. He argues current LLMs are not yet sufficient for fully autonomous trading due to gaps in domain knowledge, data access, execution, and security, and introduces Conductor as an execution harness built on Hummingbot to address some of these practical issues. He closes by outlining a multi-stage hackathon with exchange partners to experiment with trading agents and gather feedback from teams.

Key takeaways

  • The presenter positions Hummingbot as a widely used open-source framework for crypto algorithmic trading and market making, with substantial reported community trading activity across many venues.
  • He maps a four-step coding agent loop onto trading, stressing the need for testable strategies, robust data pipelines, careful backtesting, and incremental live deployment with monitoring and iteration.
  • He argues LLMs have limitations for autonomous trading and introduces Conductor, built on Hummingbot, along with a trading-agent hackathon to explore execution, data, and security challenges in practice.