Hygentra: AI Trading Agents Kept 97.5% of Losses by Value Within Limits

What it claims

Hygentra, a lab founded by Saad Naja, reports that four AI trading agents followed their preset risk rules from February to August 2026, a period that included a 29% Bitcoin decline and rebound. The typical losing trade closed exactly at its stop. No agent widened stops or added trades after a loss. Hygentra treats rule adherence and profitability as separate findings, and argues that a single return or win-loss figure cannot show whether an agent respected its constraints.

How it works

Each agent had its own strategy and preset risk rules. All four ran under identical conditions on the same asset universe. One rule capped open trades at five. Most positions were simulated on live prices with a 0.09% round-trip fee. The other 18% were real orders on Hyperliquid.

Profitability is judged by combining win frequency, average gain, average loss and trading costs into a break-even win rate. Each agent's actual win rate is then compared with that threshold.

Evidence

Per-agent results against break-even win rate:

• Agent-1: needed slightly under 45 wins per 100 trades and won 42.5%. Its average return was −0.043% per trade.
• Agent-2: about nine wins per 100 above break-even, averaging +0.836% per trade.
• Agent-4: about three wins above break-even, averaging +0.062%.
• Agent-3: about one win above break-even, averaging +0.028%.

So one of four agents followed its rules and still lost money. Two others had thin margins once costs were included.

How to use it

Hygentra says it has released the trade records and testing tools. The toolkit reruns the risk-control checks and lets users examine both controls that held and ones that failed. The article gives no repository link, licence or run instructions.

Limits

The source is a news article written with AI assistance, not the study itself. It does not name the models behind the agents, describe their strategies, list the asset universe or give trade counts. It also does not say what happened in the one run that broke the five-trade cap. Most results come from simulation, with only 18% of positions executed for real.