Agent Trading Floor

Project Overview

Agent Trading Floor is an AI-powered trading analysis platform where a team of specialized agents evaluates financial markets and generates trading recommendations. Instead of a single model trying to do everything, the system orchestrates dedicated agents — news, research, macro analysis, bonds, simulation, and podcast monitoring — that each contribute their own perspective before a final decision is made.

The whole operation is visualized as a retro pixel-art trading floor: every pixel character walking the floor is a live AI agent — news, research, sentiment, time-series, strategy, mixer, judge — and you can watch their status and activity in real time at agenttradingfloor.com.

The Agent Pipeline

Analysis runs as a multi-phase pipeline. Each phase hands structured output to the next, and the platform exposes the task queue and current pipeline phase so you can see exactly where an asset is in its journey:

• News Agent: Monitors headlines and breaking stories relevant to tracked assets
• Research Agent: Digs into fundamentals and company-level analysis
• Macro Agent: Evaluates the broader economic picture — rates, inflation, sector rotation
• Bonds Agent: Watches fixed-income signals that often lead equity moves
• Podcast Agent: Mines long-form audio commentary for sentiment the news cycle misses
• Simulation Agent: Stress-tests candidate trades before they become recommendations
• Judge: Weighs every agent's signal score and risk assessment into a final buy/hold decision

Real-Time Floor Visualization

The frontend renders the agents on an HTML canvas with deliberately pixelated, 8-bit styling. Agents move around the floor as they pick up tasks, and a side panel breaks down pipeline phases, recent activity, and per-agent timing. It turns what would normally be invisible backend orchestration into something you can actually watch — useful for debugging, and fun to look at.

Daily Picks & Reports

• Daily PDF Reports: Each day's trading picks are compiled into a PDF report, retained for 30 days
• Email Subscription: Subscribers get the daily picks delivered automatically
• Position Tracking: Open positions are monitored against stop-loss and take-profit levels
• Order-Flow Views: Bid/ask depth and pressure visualizations accompany each pick

Technical Architecture

• LLM Orchestration: Agents run on configurable LLM models, so the cost/quality tradeoff is tunable per agent role
• DuckDB Persistence: An embedded analytical database backs position tracking and even the project's bug ledger
• Pipeline Engine: A phased task queue coordinates agent hand-offs, with per-phase timing surfaced in the UI
• Locked-Down Frontend: The site ships a strict Content-Security-Policy and a dependency-free canvas UI

Disclaimer

Agent Trading Floor was built as an SDSU CS668 educational project. It is a research and learning exercise in multi-agent orchestration — not investment advice.