The typical losing trade closed exactly at its stop, Hygentra reported. The audit also checked position limits: in 23 of 24 trading runs, the agents respected the restriction of no more than five trades open simultaneously.
Most positions were simulated using live prices, with a 0.09% round-trip fee included. The remaining 18% were real orders on Hyperliquid.
Losses did not trigger bigger bets
No agent increased its risk or trading frequency after losing, according to Hygentra. They did not widen stops following setbacks or place extra trades to recover losses—behaviors the research contrasts with human overconfidence and revenge trading.
Rule adherence and profitability were separate findings. Hygentra argues that a single return or win-loss figure cannot establish that an agent respected its constraints. Its analysis instead examines winning frequency, average gains, average losses and trading costs together.
Agent-1 illustrates the distinction. Given its gains, losses and costs, it needed slightly fewer than 45 wins per 100 trades to break even. Its win rate was 42.5%, and its average return was −0.043% per trade.
The other three agents exceeded their break-even win rates. Agent-2 recorded about nine additional wins per 100 trades above the required level, averaging +0.836% per trade. Agent-4 exceeded its threshold by about three wins, averaging +0.062%, while Agent-3 exceeded its threshold by one, averaging +0.028%.
Published records put compliance under scrutiny
Hygentra has released the trade records and testing tools so brokers, investors and regulators can independently check the agents’ conduct against their assigned rules. The toolkit supports rerunning risk-control checks and examining both successful controls and failures.
That approach differs from public leaderboards that rank agents on a single headline ratio without showing compliance, the lab said. Its research focuses on auditing what deployed agents actually did, rather than scoring only their results.
“Agent autonomy should be earned, not assumed. We’ve published tests so anyone can rerun them before allocating capital,” said Saad Naja, Hygentra’s founder and CEO.
Naja, a former product leader at Goldman Sachs, Deutsche Bank and Kraken, also called for “common stress tests and public failure reports, with clear rules on who answers when safeguards fail,” so customers can compare trading products using evidence.
Institutional trading discipline for retail investors
Hygentra says its findings demonstrate that AI trading agents can automatically enforce institutional risk controls while producing records others can verify. The retail opportunity it identifies combines access to a professional trading process with the ability to check that the process was followed.
The study describes institutional controls that check compliance before orders, impose hard risk limits and block allocation or concentration breaches rather than flagging them afterward. Hygentra presents programming those controls into agents as a way to make that discipline available beyond institutional desks.
For retail investors, the lab’s proposed benefit is not simply automated execution. Its published trade record and evaluation toolkit provide a way to inspect adherence to preset rules alongside the returns each strategy produced.
Article produced with the assistance of artificial intelligence and reviewed by the editorial team.
