Research.

Our goal

Our goal is to use AI agents to build funds that outperform the S&P 500. Not tips, not signals: three real, fully managed portfolios you can watch, with the reasoning behind every decision published the same day.

How TradingAgents works

TradingAgents is open research that runs a small team of AI agents the way a real investment desk works. Instead of asking one model to guess, each agent has a job, they argue it out, and a manager makes the final call.

• News agent — reads the headlines and earnings and works out what actually matters.
• Macro agent — watches interest rates, inflation and the wider market. The weather overhead.
• Sentiment agent — reads the mood. Is the crowd greedy, scared, or half asleep.
• The researchers — a bull and a bear take opposite sides and genuinely argue each company.
• The trader — turns that argument into a concrete move.
• The risk team — stress-tests the move before it goes anywhere near real money.
• The portfolio manager — weighs the lot and sets the final position.

The point of the arguing is simple. Ask one model "should I buy this?" and you get a confident, shallow yes. Make a bull and a bear fight over it, then make a risk team sign off, and the weak spots come straight to the surface.

How we use it

Most people who pick up TradingAgents point it at a single stock and ask whether to buy. We do something different. We funnel that research into three managed funds, so the agents are not just picking one name, they are running whole portfolios.

Every trading day at market close, the full team runs again: the news, macro and sentiment agents gather what happened, the bull and bear argue it out, the risk team checks it against each fund's mandate, and the portfolio manager rebalances. Nothing sits still by accident. Every position is re-examined, and whatever changes get written up in plain English on the Commentary page that same day.

The three funds

• Whitewater — our high-risk fund. The agents take real swings here, backing fast-moving companies in space, AI and frontier tech. Bigger upside, bumpier ride.
• Tidewater — the balanced middle. Broad index exposure and quality companies, aiming to beat the market without the stomach-drops.
• Stillwater — the defensive fund. Capital preservation first: index funds, bonds, gold and steady names. Deliberately boring.

Same engine, same daily process, three different appetites for risk. See how each one is performing.

How a call actually gets made

Every position starts as a five-step relay. Nothing skips a stage, and each one can send the decision back down the line.

1. 01
   Gather. The analyst agents pull the day's prices, fundamentals, recent news and the macro backdrop for each name on the watchlist.
2. 02
   Debate. A bull and a bear researcher argue the case back and forth, for as many rounds as it takes to stop repeating themselves.
3. 03
   Propose. The trader takes the stronger side of that debate and writes a concrete move: buy, hold or sell, and how much.
4. 04
   Stress-test. The risk team attacks the proposal on a risk-first basis, and can send it back if it runs too hot for the fund's mandate.
5. 05
   Commit. The portfolio manager makes the final call, sizes the position, and writes the one-line reason you read on the Commentary page.

Under the hood

The heavier version: TradingAgents is built on LangGraph, which wires the agents into a graph they pass their findings through. It pulls live prices and fundamentals, recent news and macro data, runs each agent as its own language-model call, and lets the researcher and risk stages run several rounds of back-and-forth before the manager commits. It is open research from Tauric Research, so anyone can check exactly how it reaches a call: read the paper on arXiv or the code on GitHub.

Who is behind it

LiquidAssets is run by a finance and computer science student who has traded the US and New Zealand markets for years. The academic work does the reasoning. The judgement about how to point it, and at what risk, is ours.

Model portfolios only. Not financial advice. No real trades are placed by this prototype.