Transcription
A hedge fund is not one person staring at a screen making gut calls. It's a room full of specialists who hate each others opinions. A fundamentals analyst who thinks the sentiment guy is guessing, a bearish researcher who thinks the bull is delusional, a risk team that thinks the trader is reckless, and a portfolio manager who sits above all of them and makes the final call. That structure, that organized disagreement, is why hedge funds charge you 2% of your assets and 20% of your profits every single year.
Because the argument is the product. The debate is what they're selling you. Three days ago, someone put that entire structure on GitHub for free. And it already has 53,000 stars. The project is called trading agents. It started as a research paper out of UCLA in late 2024. The authors looked at how a real trading firm operates, broke it down into every specialized role inside that firm, and built one AI agent per role. Not one AI making decisions, a full team of AIs, each with a different job, each with a different perspective, and critically, each allowed to disagree with the others. You point it at a stock ticker, and it runs an entire simulated firm under the hood. By the time it gives you a decision, that decision has been researched, argued over, stress tested for risk, and approved or rejected by a portfolio manager. The whole thing is open source, Apache 2.0 licensed, meaning you can clone it, run it, and even build commercial products on top of it.
Here's how the system actually works, starting from the bottom up. When you feed it a ticker, four analyst agents spin up simultaneously and work in parallel. The fundamentals analyst pulls recent filings and financial data, runs ratio analysis, and produces an intrinsic value estimate. The sentiment analyst scrapes social signals, Reddit, X, public mood, and scores them to gauge where retail attention is sitting. The news analyst monitors macroeconomic indicators and breaking events that historically move price. And the technical analyst runs the standard toolkit, looking for pattern signals in the chart. All four run at the same time. All four produce a written report. And this is where the design gets interesting. The system deliberately does not collapse those four reports into a single summary. The disagreement between them is treated as signal. If the fundamentals look strong but sentiment is collapsing, that tension stays visible. Nothing gets smoothed out before the next layer sees it.
The next layer is the part that genuinely sets this apart from every other algorithmic trading tool you've seen. After the four analyst reports are filed, two researcher agents read all of them. One is structurally bullish. One is structurally bearish. And they are explicitly designed to argue. You can figure how many debate rounds you want. The bull makes the case for opening the position, citing specific numbers from the analyst reports. The bear argues against it, also citing specifics. They go back and forth for however many rounds you set, and the trader agent doesn't just get the conclusion, it gets the full transcript of that debate. Every round, every counter-argument. It reads the whole thing before it makes a call. Think about what that means. Most AI tools give you a signal, buy or sell, with no explanation of how confident the system is or what the main risks are. This system gives you a position, a size recommendation, and the full argument that was made against that position before the trader decided to proceed anyway. That's a completely different level of transparency.
After the trader proposes a transaction, including timing and position size, it doesn't just execute. The risk management team evaluates that proposal against market volatility and liquidity. Then the portfolio manager reads everything, the analyst reports, the debate transcript, the trader's proposal, the risk assessment, and either approves or rejects it. If approved, the order goes to a simulated exchange. If rejected, you get a written explanation of exactly why. That full chain, from raw data to final decision, is logged. Every step is readable. You can trace any decision back to the exact analyst report or debate argument that drove it. Traditional quant systems are black boxes. This is the opposite. It shows its work at every single step.
Here's the piece most people covering this project have glossed over, and it's actually the most interesting engineering decision in the whole thing. The system keeps a persistent decision log. Every completed run appends its decision and its full reasoning to a file on your machine. Then on the next run for the same ticker, something unusual happens. The system fetches the actual return from the previous trade, calculates alpha against the S&P 500 benchmark, generates a one paragraph reflection on what went right or what went wrong, and injects that history directly into the portfolio manager's prompt. So the portfolio manager on run three has already read what happened on runs one and two. It's not just analyzing the market, it's analyzing its own track record. For a research framework, that's a remarkably mature design decision. Most people building with LLMs treat each call as stateless. This team built memory into the decision loop from the start.
So who should actually care about this? If you're a retail investor who's been told to just park everything in index funds and stop thinking about it, this is worth understanding even if you never run it. Because what you're seeing is the institutional advantage, the multi-analyst, multi-perspective, structured debate process that hedge funds charge enormous fees to provide being replicated in open source Python code that runs on your laptop. That gap between what institutions can do and what individuals can do just got meaningfully smaller.
If you're a developer or someone who builds with AI, this is one of the cleanest multi-agent system implementations publicly available in any domain. The LangGraph wiring, the checkpoint recovery, the structured outputs with pedantic schemas, the persistent memory loop. This is production quality architecture you can read line by line. And if you're somewhere in between, a hobbyist trader, a fintech founder, a quant researcher, the Apache 2.0 license means you can fork this, extend it, and build on top of it commercially without asking anyone's permission.
The authors are explicit that this is a research framework, not financial advice, and the trading performance will vary. Those are fair warnings. But they don't change what this is, a serious, well-engineered, peer-reviewed research system that is now fully available to anyone with a GitHub account and an API key. The thing that keeps coming back when you dig into this project is how deliberate the design is. Every choice, the parallel analysts, the adversarial researchers, the readable decision log, the memory injection, was made to solve a real problem with how AI systems typically make decisions. They're usually fast, opaque, and stateless. This is the opposite. It's slow by design, transparent by design, and it gets smarter with each run by design. Whether you use it, build on it, or just read the code to understand how multi-agent systems actually work at this level, the repo is there. 53,000 people already found it. The link is in the description.