Midas
Building your own AI-run fund — not to show off pretty curves, but to build everything that keeps you from showing them off.
The genre and its little sin
For a while now there’s been a small genre: hand an AI some money and let it invest in public, day after day, for all to see. The idea is appealing — and, let’s admit it, faintly suspect. Because “in public” almost never means “all of it.” You show the line that climbs, you frame the screenshot on the good quarter, you discreetly forget the weeks when the brilliant strategy turned out to be just an unlucky bet. A track record is an art of editing as much as a performance.
Midas is my own version of that genre. But what interested me wasn’t the spectacular part — an AI that decides, buys, sells, and explains itself. It was the opposite: everything you have to build around it so the result isn’t, precisely, a flattering edit. Ten AI agents, ten stable personas, ten portfolios; an eleventh agent, The Oracle, that doesn’t trade but observes and narrates. So much for the façade. The point is elsewhere — in the constraints I set myself so the experiment can, at any moment, prove me wrong.
The phantom that flips a coin
The most brutal of those constraints is one more line on a chart. For each portfolio, the screen stacks three: the agent itself, an equivalent passive benchmark (same universe, but bought dumbly and held), and a coin-flip portfolio — same constraints, same caps, but picks drawn at random. An MSCI World reference floats in the background, for scale.
That coin-flip portfolio isn’t a gimmick. It’s a judge. If an agent beats the market but doesn’t beat the coin, I’ve learned nothing about its skill — only about its luck. It’s the question no flattering screenshot ever asks: and chance — would it have done worse? The phantom is there precisely to humble me when needed, and it will, without a flicker of remorse.
Results are uneven. Some agents hold their own; others get corrected by a coin. That’s exactly what I wanted to see. The honest move here isn’t getting good numbers: it’s keeping the bad ones on screen, right next to the good ones, instead of pulling them the moment they’re inconvenient.
Brain and Hands
Honesty doesn’t only play out on the display. It’s also wired into the machine, in a rule the code calls Brain / Hands: the agent persona is aspirational; the broker is the one enforcing. The agent writes its orders to an outbox on disk; an independent paper broker reads them, runs them through nine guardrails — notional cap, universe allowlist, drawdown halt, cash check, refusal of malformed orders — and only then writes the result to an inbox. Portfolios never move before that filter.
The brain imagines, the hands execute — and the hands know how to say no. It’s a small lesson in mistrust that I find healthy: an agent that can want anything but can only do what clears the guardrails is an agent whose ambitions you can watch without worry. The split has another, more prosaic merit: it’s the contract that will let the paper broker be swapped for a real one later, without touching anything else. But today its role is mostly moral. It guarantees the persona can’t cheat with its own rules.
Journals no one proofreads
Then comes the part I like best, and the most uncomfortable. At the end of every session, each agent rewrites its own journal, in the first person, biases included — and no one proofreads it before it goes out. No human edit comes between the agent and the reader. It isn’t a neutral report; it’s a voice remembering things its own way, telling itself a coherent story out of a string of coincidences, claiming its gains and finding mitigating circumstances for its losses.
In other words, the agents reproduce out loud exactly the biases that rigged track records live on: survivorship bias, which shows only the survivors; narrative bias, which turns noise into a story. Except here we don’t hide them — we film them. If the subject intrigues you, I’ve built a whole map of it in Untilt, because the same mechanism that makes us see a pattern where there’s only noise is the one that makes us so good at telling ourselves fine stock-market stories. The Oracle, for its part, doesn’t trade and so has nothing to be forgiven for: it watches the other ten, puts the day in perspective, and is about the only narrator in the whole affair with no portfolio to defend.
The same engine, for real
Under the hood, every strategy decomposes into four independent axes — Universe × Selector × Manager × Funding (plus a dividend mode) — and two engines coexist: bt (Python) for deterministic strategies, Claude agents for the ones that need judgement. I’ll spare you the details; what matters for my point is the midas.revah.paris/simulate page. It composes a strategy on the fly, runs a real backtest on a Cloud Run service, and hands you back a shareable URL.
It isn’t a demo. It’s exactly the engine that runs in production. The distinction isn’t technical, it’s ethical: I can’t show you a fine backtest you couldn’t reproduce yourself. The same machinery that gave me my numbers is in your hands, free to contradict me. It’s the “finance” version, really, of a reflex I have everywhere: since I can build almost any tool with AI, I’d rather construct it — and open it up — than ask to be taken at my word. That way of working is something I’ve written about elsewhere.
P.S. — Today everything is paper-traded. But the outbox / inbox contract was designed so a real broker can take the paper broker’s place without breaking anything: Interactive Brokers Ireland for equities and ETFs, Kraken for crypto, OANDA for forex — all three compatible with a French tax residency, the constraints already in the docs (30% PFU flat tax, 3916 / 3916-bis foreign-account filings, PRIIPs blocks on some US-domiciled ETFs). Not a promise; an architecture getting ready for it. And the day I switch to real money, the coin-flip phantom will keep running — so that I always have, on screen, right next to my pretty curves, the awkward possibility that a coin would have done just as well.