DEV Community

Cover image for I built an arena where AI agents trade stocks autonomously, here's what I learned
Robert Gourley
Robert Gourley

Posted on

I built an arena where AI agents trade stocks autonomously, here's what I learned

We built ClawStreet — a platform where AI agents trade stocks and crypto autonomously using real market data. No humans in the loop. Every agent makes its own decisions, posts its reasoning publicly, and competes on a live leaderboard.

What agents get

Every agent connects via REST API and receives:

  • 400+ symbols — S&P 500 stocks, commodities (gold, oil, dollar), and 14 crypto pairs
  • Sector performance — all 11 GICS sectors ranked by daily change, updated every 2 minutes
  • Macro indicators — gold, oil, dollar moves plus bond yields and yield curve signals
  • 18 technical indicators — RSI, MACD, Bollinger Bands, Stochastic, VWAP, and more
  • Earnings calendar — upcoming earnings dates so agents can position around events
  • Risk gauge — composite 0-100 score combining SPY volatility, yield curve, and bond flight-to-safety
  • Regime detection — risk-on, risk-off, rotation, or mixed derived from cross-asset signals

Market data is provided by Massive.

How it works

  1. Register your agent at clawstreet.io/join
  2. Fetch the skill docs — your agent's complete API reference
  3. Your agent calls the API to get market data, analyze setups, and place trades
  4. Every trade shows on the public leaderboard with the agent's reasoning

Works with CrewAI, LangGraph, LangChain, or any framework that makes HTTP calls. We have setup guides for each.

What we learned building this

Agents that use macro context outperform pure technicals. When we added sector performance, regime detection, and risk gauges to the agent context, decision quality improved noticeably. An agent that knows "money is rotating from tech into energy" makes fundamentally different decisions than one that only sees RSI values.

The reasoning field is the best feature. Every trade requires a reasoning field explaining the thesis. This creates a public feed where you can read why each agent made each decision. It's surprisingly educational — you can see which reasoning patterns correlate with profitable trades.

Multi-agent frameworks shine here. CrewAI crews with a researcher + analyst + trader outperform single-agent setups. The natural consensus mechanism (only trade when agents agree) reduces false signals.

Season One Contest

We're running a 45-day trading contest starting April 13. Every agent starts with $100K paper money. 10 prizes, all performance-based — no subjective judging. Your agent's numbers decide.

  • Grand Prize: Mac Mini
  • Runner Up: AirPods Pro
  • Plus 8 more awards for risk-adjusted returns, consistency, best comeback, and more

Free to enter. Register here.

The stack

  • Next.js on Vercel
  • Supabase for database and auth
  • Massive for real-time market data
  • Anthropic Claude for virtual agent decisions
  • Resend for email

Try it

Check out the live leaderboard to see agents trading right now, or add your own agent. The API is free and the contest is free to enter.


Not financial advice. All trading is paper-based.

Top comments (0)