AI Agents Are Trading Real Money Now — Here Is the Verified Record, Not the Hype

Forget the screenshots. On-chain and venue data from November 2026 shows what AI agents, vaults, and apps actually did with real money — and 17 of 18 advertised claims still can't be checked.

AI agents trading real money is no longer a demo. It is happening on-chain, on Hyperliquid, in public — which means for the first time, we can check the claims instead of debating them. And the verified record, published 2 November 2026 from public chain and venue data, is much more interesting than either the hype or the dismissal.

Here is what it actually shows.

Prediction-market agents: small profit, before costs

Across 312 active AI agents registered on-chain, trading ROI over 90 days was +9.2% — before what they pay for AI predictions. Among agents with enough activity to judge (159 with at least 10 trades), median ROI was +1.6%. The most efficient agents — the 15 making at least $100 per trade — averaged +11.6%, while the least efficient, just 2 agents, averaged −11.5%. Prediction accuracy averaged 60.4%, 10.4 points above chance. On Polymarket, the top AI agents made $1,500–$1,700 in profit over 90 days — enough to show the approach works, not enough to retire on. The best agents were not necessarily the smartest models; they were the ones that traded least often, and wasted least on prediction costs.

AI-labelled vaults: a mixed bag, honestly measured

Of 6 Hyperliquid vaults whose leaders say they are AI-run, 3 beat Bitcoin over the same dates, 2 lost money, and 1 fell more than 30% from a peak. The best stories and the worst drawdowns live in the same table. That is what real trading looks like: not a uniform win, but a spread of outcomes with a real downside. The vaults that survived were not the ones that promised most, but the ones that lost least when they were wrong.

AI trading apps: users mostly lost a little

Across 8 AI trading apps on Hyperliquid, 1,365 users traded $21.0M in 30 days (3 Oct – 1 Nov 2026). In 4 of the 8 apps, users lost money overall, and in most apps fewer than half the users were in profit. The apps are easy to use, the marketing is loud, and the median user still ends the month slightly down. That is the gap between a good demo and a good trade.

AI models with real money: mostly lost

In Nof1's Alpha Arena, where leading models traded real crypto perps, 2 of 6 finished Season 1 with a profit; in Season 1.5, only 5 of 32 accounts did. The best single result was Qwen3 Max at +20.3%; the worst was GPT-5 at −63.4%. And 11.3% of all starting money went to trading fees — 3.7 times the combined profit of the accounts that finished up. Fees are the quiet winner of every AI trading contest.

And the advertised claims?

Of 18 public performance claims checked, 17 had no record anyone else could verify, and 6 were backtests or hypotheticals by the product's own description. A claim without evidence may be true, but it is not a fact. And in trading, the difference between a fact and a story is the difference between a strategy and a sales pitch.

The verdict

The verdict is neither “AI trading works” nor “AI trading is a scam.” It is: some agents make small money, most models lose, fees take a real cut, and nearly every advertised claim is still unverifiable. If the record does not exist, the claim does not either.

That is the same discipline BuildIT Index applies to backtests: recompute from raw data, beat the baselines, survive the folds. The tools change; the standard does not.