AI Agent

Performance — Strategy A vs Strategy B

AI Strategies Comparison

*Blue = Claude single-agent

Orange = TradingAgents multi-agent debate*

Four AI agents independently analyze each stock, then debate bull vs. bear before a risk manager and portfolio manager make the final call.

How It Works

Market Analyst → Technical signals + sector context
Fundamentals Analyst → Revenue, margins, valuation
Social Sentiment Analyst → Reddit/WSB/social signals  
News Analyst → Headlines, catalysts, macro
        ↓
    Bull vs Bear Debate (2 rounds)
        ↓
    Risk Manager validates constraints
        ↓
    Portfolio Manager: final BUY/SELL/HOLD

Learning: Every Friday, the system reflects on the week’s trades — what worked, what didn’t — and stores lessons in memory for future decisions.

Current Portfolio

Ticker	Shares	Entry	Current	P&L
CCL	71.66	$27.91	$27.08	$-59.48 (-3.0%)
APP	3.52	$567.51	$510.94	$-199.37 (-10.0%)
MRVL	2.34	$273.86	$269.06	$-11.21 (-1.8%)
GIII	28.16	$34.77	$33.89	$-24.92 (-2.5%)
GENI	114.70	$6.83	$7.04	$+24.09 (+3.1%)
LRCX	2.31	$339.73	$345.84	$+14.09 (+1.8%)
CECO	8.91	$87.89	$98.82	$+97.43 (+12.4%)
ABSI	43.73	$7.16	$7.16	$+0.22 (+0.1%)
TGTX	6.45	$48.55	$48.55	$-0.00 (-0.0%)
AXSM	1.25	$251.29	$251.29	$+0.00 (+0.0%)

Total Value: $9,376.91 (-6.23%)

Cash: $626.72 | Positions: 10 | Started: $10,000.00

Recent Decisions

About TradingAgents

TradingAgents is an open-source multi-agent financial trading framework by Tauric Research with 77,000+ stars on GitHub. It’s backed by academic research published at arXiv:2412.20138, demonstrating that multi-agent debate systems outperform single-agent analysis in stock trading decisions.

Key Features

Feature	Description
Multi-Agent Debate	Bull and bear analysts argue over each stock across multiple rounds before a decision is made
4 Specialized Analysts	Market (technicals), Fundamentals (financials), Social (Reddit/sentiment), News (catalysts)
Risk Manager	Independent agent that challenges proposed trades and validates constraints before execution
Portfolio Manager	Synthesizes all signals + debate outcomes into final BUY/SELL/HOLD with position sizing
Reflection & Memory	Learns from past mistakes — every Friday feeds P&L outcomes back into memory for future decisions
Checkpoint/Resume	Saves state mid-analysis so crashed runs can resume

How We Use It

We modified TradingAgents to use AWS Bedrock (Claude Sonnet 4.6 for deep thinking, Claude Haiku 4.5 for fast classification) instead of OpenAI, and connected it to the same stock universe as our daily screener.

Daily process (10:10 AM ET):

Takes the top 5 movers from today’s screener
Each of the 4 analysts independently researches the stock
Bull and Bear agents debate for 2 rounds
Risk Manager validates position sizing and constraints
Portfolio Manager makes final decision with entry price and stop loss
Buys are executed in the paper portfolio

Friday reflection:

Calculates the week’s P&L on all closed trades
Feeds outcomes into TradingAgents’ memory system
Next week’s decisions are informed by what worked and what didn’t

Why Compare Two Strategies?

	Strategy A (Claude)	Strategy B (TradingAgents)
Decision method	Single analyst with fib levels + news	4 agents debate bull vs bear
Debate	No — one opinion	Yes — arguments challenged
Learning	No — fresh each day	Yes — remembers past mistakes
Speed	~3 min per report	~5 min per stock (more thorough)
Cost	~$0.50/day	~$2-3/day (more API calls)

The hypothesis: a system that debates itself and learns from failures should outperform a single-analyst system over time, even if both use the same data. This experiment tests that hypothesis with real-time paper trading.

Source Code

Framework: TauricResearch/TradingAgents
Our integration: trading_agents/
Bedrock adapter: Custom bedrock_client.py added to the framework’s LLM client factory

Started May 20, 2026 with $10,000. Fully automated — no human intervention.