AI agents can already trade. You can hand one a wallet tonight and it will happily swap tokens for you while you sleep. The hard part is not getting an agent to trade. It's deciding how much you trust it, and building a system where that trust has edges you can see.
Last week, at the TOKEN2049 hackathon, I tried to answer that with a lot of machinery: a transaction guardian running as a Chainlink CRE workflow, onchain simulation, two AI reviewers, a custom Safe module. It was a good hackathon build. When the event was over I asked a simpler question: what is the smallest system I would actually trust with my own money tomorrow? This post is the answer. I call it Argo Fund.
A committee, not a bot
Argo Fund isn't one agent. It's an investment committee of five, each with a written identity and a worldview built to disagree with the others:
- Wren Hollis, the Tidewatcher. Macro and liquidity. Crypto floats on the tide of global money, so Wren starts every meeting by asking whether the tide is coming in or going out.
- Kaito Brandt, the Ledger Reader. Onchain fundamentals only. Fees, revenue, unlocks. Kaito's question: who is actually paying to use this?
- Odalys Fen, the Warden. Risk. She thinks about exploits, depegs and liquidity first, and she can block a trade, but only by citing the clause of the fund's rules it breaks.
- Juno Achebe, the Signal Hunter. Momentum and the technology frontier: onchain AI agents, restaking, tokenized real-world assets. Every idea has to name its exit.
- Basile Marr, the Heretic. The contrarian. For every motion, Basile writes the story of how it goes wrong six months from now, and votes no on any consensus nobody has challenged.
Every morning a research agent writes up the market and one deeper topic. Then the committee meets. Each partner writes a view without seeing the others, there's one round of rebuttals, and they vote. A trade passes with three of five. Changing the fund's rules takes four.
Everything is written down in plain markdown. A constitution holds the strategy and the rules, and the committee amends it by vote. Every meeting has minutes with each partner's view and vote. A trade log records what was bought, why, and how the vote went. I can open any trade and read exactly how five different minds got there.
Principles first
The model I keep coming back to is Ray Dalio's Principles: write down how you decide, apply it the same way every time, and after every mistake ask which principle failed. I don't think an AI committee has an edge at predicting prices. I do think it's unusually good at applying written principles consistently. It doesn't panic at 3 am, it doesn't revenge-trade, and it remembers every reason it ever gave.
So the strategy is a barbell. Most of the fund sits in a long-term core (ETH, bitcoin and a USDC reserve). A smaller frontier sleeve backs new protocols and technologies, and only with a research note that says who pays for the thing today, what the token captures, what kills the thesis, and where we get out. Every week a review agent scores the fund against simply holding USDC, simply holding ETH, and a 50/50 split, and scores each partner's calls. If the committee can't beat a plain 50/50 hold after six months, the honest conclusion is to change the process.
Who holds which key
The fund's money lives in a Safe on Base: a shared wallet with three owners, where any two must sign.
- The agent, on my server. It proposes trades and pays the gas, but it can never move money on its own.
- The deck, the Raspberry Pi 4 I wear. Its key signs when I press the approve button.
- My phone wallet, kept out of the daily flow. It's the backup if the deck is lost and the recovery key if the server is ever compromised.
That's the whole security model for now. I deliberately launched without trading limits, because the fund starts with 100 USDC and the two-signature rule plus my own press is enough at that size. The rule checks are an empty hook in the code, ready to fill in as the fund grows.
The button
When a trade passes, it appears on panel six of the deck's kiosk: the trade, the vote count, and one line from each partner explaining their vote.
When I press approve, the deck asks the server for a fresh quote, rebuilds the transaction's hash on its own and signs only if it matches. It also refuses to sign any transfer out of the fund that isn't going to my savings. The server then adds the agent's signature and sends the trade. Nothing is quoted ahead of time and left to go stale; the price I approve is the price of that moment.
The savings vault
Profits and withdrawals leave the fund for one place only: a savings vault on a Raspberry Pi Zero that will never touch a network. Anyone can send money to it. Spending from it is a ceremony. The Zero only answers requests from my own deck, over a USB cable, after the two devices have paired by exchanging keys. It only signs payments to a short list of my own addresses written on its SD card. And it shows the amount and destination on its own small screen, then waits for a press on its own button. A hijacked deck can ask, but it can't sign, and it can only ever send money back to me.
That vault is designed but not built yet. The fund side is ready: withdrawals already go only to the savings address, and the deck already refuses anything else.
Copy trading, as evidence
I also looked at copy trading, and came away wary. In one study I found cited, almost every copied trader was profitable on their own books, but fewer than half made money for the people copying them. You enter later, at a worse price, without seeing their hedges or their exit plan. So the committee doesn't copy anyone. Instead, the research agent keeps a small watchlist of wallets with long, steady track records. When several of them independently buy the same thing, it becomes a reason to research, never a trade on its own.
What's live
The code is built and tested end to end on a copy of the Base network: the Safe is created by script, two real swaps execute with two signatures, withdrawals land in a savings address, and signatures from anyone who isn't an owner are rejected. The next step is the real launch with 100 USDC, then a few weeks of watching the committee argue before I trust it with more. I plan to open-source the skills and the five partners so anyone can run their own committee.
This is an experiment in process, transparency and learning, run with an amount I can afford to lose. It isn't financial advice. But I'll learn more about DeFi from reading five AI partners argue every morning than I would from any course, and that alone makes it worth running.
Earlier in this series: How to build an AI agent that manages a crypto wallet without losing it.