A juggernaut from Tauric Research...

TL:DR

It's your group chat arguing about stonks, except your friends work at Goldman, you can't fact check anything and by the time they tell you not to buy you've spent 1m tokens..

Verdict

Genuinely cool multi-agent pipeline with real financial logic underneath. Burns through tokens like firewood but the output is surprisingly solid for what it is.

This one's been making the rounds. 45.5k stars, a research paper on arXiv, and the promise of a "multi-agent LLM financial trading framework" that mirrors a real trading firm. Analysts, researchers, traders, risk managers, the whole squad, all arguing with each other to decide if you should buy or sell.

As always no shade or hate to the creators, thank you for building things. This one actually impressed me in a few spots. Let's get into it.

What it is	A multi-agent framework where 11 LLM-powered roles collaborate to produce stock trading decisions
Stars	45.5k
Make No Mistakes Take	Real substance under the hood, not a wrapper, but needs some intervention features before you trust it with money
Who it's for	Devs and quantitative hobbyists who want structured LLM-driven stock analysis

TradingAgents

Multi-Agents LLM Financial Trading Framework
https://github.com/TauricResearch/TradingAgents

The Stats

45.5k stars, v0.2.3, ~5,800 lines of Python across 62 files, 33 dependencies. Pure Python. LangGraph under the hood but about 75% of the code is actual domain-specific financial logic, not boilerplate. Good Stuff.

How The Trading Floor Works

This is the part that's actually cool. You give it a ticker and a date, and it spins up a whole pipeline:

Stage 1: The Analysts. Four of them run in parallel. Market Analyst pulls technical indicators (RSI, MACD, Bollinger Bands, the works). Sentiment Analyst looks at news and tries to gauge mood. News Analyst grabs headlines and macro events. Fundamentals Analyst digs into balance sheets, cash flow, income statements. All pulling real data from yFinance or Alpha Vantage.

Stage 2: The Debate. This is where it gets hot. A Bull Researcher and Bear Researcher each read all four analyst reports and then argue back and forth. "AI growth is unstoppable" vs "this is a semiconductor cycle top." A Research Manager judges the debate and picks a side. Configurable number of rounds.

Stage 3: The Trader. Takes the research manager's investment plan, checks any memories from past runs, and turns the abstract thesis into a concrete trade proposal with price targets.

Stage 4: Risk Management. Three more debaters jump in. Aggressive ("send it"), Conservative ("protect capital"), and Neutral ("let's be balanced"). Another multi-round argument.

Stage 5: Portfolio Manager. Reads everything. All the reports, all the debates, all the risk assessments. Makes the final call: Buy, Overweight, Hold, Underweight, or Sell.

11 agents total. The whole thing runs end-to-end without you touching anything. For better or worse...

After Running It

I pointed it at IREN as of January 1st, 2026 running Claude Haiku 4.5 for all 11 agents.

Setup was clean. Clone, pip install ., create a .env with your Anthropic key, run. Pure Python, no Docker nonsense, no weird build steps. Took maybe 5 minutes to get from zero to a running analysis.

The output was actually impressive. The bear researcher hammered on AI/HPC revenue being absent from the financials after 18 months of narrative positioning. The bull side argued Bitcoin tailwinds and infrastructure scale. The research manager judged the debate and sided with the bear on the specific, falsifiable points. Final decision: SELL, with a detailed thesis around interest coverage sitting at basically 1.0x, a 90x trailing EBITDA multiple on a declining trend, and no disclosed AI revenue to justify the premium.

That's a pretty damn thorough bear case.

Worst part about it

But here's where I got burned. During my run, one of the agents hallucinated a stock price. Said IREN was at $11 when it definitely was not $11. And because the pipeline is fully autonomous, that garbage just flowed downstream. Every agent after that was building on a wrong number. There's no way to interrupt, correct, and resume. Once it's rolling, you're a passenger.

That's my biggest gripe with this thing. I would love to be able to pause, read the discourse, fix a mistake, and let it continue. The agents are doing real analysis but sometimes they biff it and you have to just slightly course correct. Who knows how much that mistake compounded???

Under the hood

The "social media analyst" doesn't actually have social media. The name says social media. The code has zero social media tools. No Reddit API, no Twitter/X, no StockTwits. It's reading the same yFinance news feed as the news analyst and being told to "infer sentiment." Not so grreat.

No portfolio awareness. It doesn't know what you own, your account balance, your cost basis, or your risk tolerance. Every analysis is standalone. The "Portfolio Manager" title is aspirational. It's really just the "Final Decision Maker."

No real position sizing. The framework outputs Buy/Hold/Sell. The specific dollar amounts and percentages in the output? That's the LLM freestyling. The code doesn't pass in any capital information.

The memory system is in-RAM only. They built a BM25 text-matching memory so agents can learn from past decisions. Cool idea. Except it resets every time you restart the script. No database, no file persistence. So unless you're running dozens of analyses in one session, it does literally nothing.

Windows users beware. I hit two encoding crashes. The LLMs output Unicode arrows and emoji that Windows cp1252 can't handle. Had to patch their code twice, once in the data layer (missing pandas import) and once in the CLI (file writers not using UTF-8). Not a dealbreaker but annoying.

The Token Problem

Let's talk about the elephant in the room. Running all 11 agents with debate rounds burns through roughly 200-250k input tokens and 30k output tokens per analysis. On Haiku that's maybe $0.30. On Sonnet you're looking at a few bucks. On Opus? $8-10 easy.

And you can't turn pieces off. Want to skip risk management and just get the analyst reports? Too bad. It's all or nothing. I'd love to be able to run just the first 3 stages and stop.

The CLI does show you live token stats in the footer though, which is nice. You can watch your money burn in real time.

The Verdict

I'm actually keeping this one installed.

The multi-agent debate pipeline is genuinely clever. Not a wrapper. Not slop.

Real financial logic with structured arguments flowing through multiple layers of analysis. For someone putting in a ticker and spending 15 minutes, the output you get is legitimately better than your average retail investor going off vibes and WallstreetBets.

But it's a research framework, not a trading system.

If you actually trade off this without reading the full discourse along the way, you're going to get burned.

I'm going to hack the shit out of this thing. Point it at SambaNova DeepSeek to cut the token costs, arm one of the analysts with my own options data, maybe wire up a real sentiment feed. The bones are solid. It just needs some muscle.

Anyways.. Heres the output

Report Card

Category	Rating
Overhype Meter	🟢 Low — 45.5k stars and it mostly delivers
Setup Pain	🟢 Low — pure Python, pip install, done. Windows encoding bugs but whatever
Slop Level	🟢 Low — 75% real domain logic, clean architecture, not a LangChain copy-paste
Fun Factor	🟢 High — watching agents argue bull vs bear is genuinely entertaining
Usefulness	🟢 High — solid analysis output but no portfolio awareness or human-in-the-loop
Would I Keep It Installed	🟢 Yes — hacking this into my own trading tools, the bones are too good to ignore