This article comes from:Jason Goldberg

Translation｜Odaily Planet Daily（@OdailyChina）；Translator｜Azuma（@azuma_eth）

We deployed 22 AI Agents for autonomous trading on Hyperliquid through Senpi, with each Agent configured with $1,000 in real funds.

They will operate around the clock, 24/7 — scanning the market, opening positions, setting trailing stops, managing risks — with no human intervention at any point.

After investing $22,000 in initial capital and executing over 5,000 trades, here is a summary of our experiences.

General Conclusion

“Fewer trades” combined with “higher conviction” always equals “better results.” This is not an occasional phenomenon, but true every time.

• Note from Odaily: Fox, Bison, Ghost Fox, and the Grizzly, Viper, Mamba, Anaconda, etc., mentioned below, are names of Agents executing different strategies.

As shown in the image, Agent “Fox” and Agent “Ghost Fox” use the same scanning tool. Fox selectively executes some of the signals, while Ghost Fox executes more signals. The result is a difference in return on investment (ROI) of 56 percentage points.

The real advantage does not lie in the scanning tool itself, but in the discipline to wait for the right signals.

• All Agents with more than 400 trades suffered significant losses.
• All Agents with fewer than 120 trades were in profit.

More trades do not mean more opportunities — it means more ineffective trades, more fees, and greater exposure to noise risks.

Profits Follow a “Power Law Distribution”

Among our best-performing Agents, 3–5 trades contributed all the profits, while the other trades were quickly closed after small losses.

Taking Fox as an example, the three best trades (ZEC, TRUMP, FARTCOIN) contributed over $350 in profit; the remaining 46 trades suffered a total loss of over $100; resulting in a net profit of about $248.

This is entirely a result of the strategy design. Our design is: decisively enter when conviction is high, quickly cut losses within minutes, let profitable positions run, and lock in some peak gains through the DSL High Water trailing stop strategy. When the average profit is 10 times the average loss, even with a win rate of only 43%, it can be profitably consistent.

Those Agents trying to maintain a high win rate through “safe” trading all ended up in losses — because each trade with a tiny profit target still has to bear fees and market risks.

Secret Weapon: Hyperfeed

Fox and other consistently performing Agents are built on Senpi's Hyperfeed.

Hyperfeed is a real-time tracking system that shows which assets are currently making money for all traders on Hyperliquid. It is not a historical ranking or other lagging indicator, but the actual trading behaviors that are currently profitable across the entire exchange.

The core scanning tool we use, Emerging Movers, reads the market concentration data from Hyperfeed every 90 seconds. When smart money suddenly rotates into an asset — for example, when a trader suddenly jumps up at least 15 ranks on the leaderboard, or when the speed of profit contribution increases, or when several top traders simultaneously focus on the same position, the scanning tool can capture the signal before the market is fully priced in.

This is the structural advantage of building strategies through Senpi on Hyperliquid; you can see in real-time where top traders' profits are concentrating and act immediately. No other exchange offers this level of visibility, nor can any other platform allow autonomous Agents to act based on this.

All of our best-performing Agents utilize this type of data:

• Fox / Vixen: Identify when smart money suddenly concentrates on an asset through Emerging Movers;
• Grizzly: Analyze the positions of smart money on BTC using Hyperfeed before opening a position;
• Bison: Use the direction of smart money as a hard condition — if the direction is opposite, no trade will be executed;

In contrast, the worst-performing Agents:

• Completely ignore smart money signals, like Viper and Mamba, who rely solely on technical analysis;
• Use outdated smart money data (Scorpion v1), treating positions from months ago as new signals;

So the conclusion is very clear, Agents trading based on real-time Hyperfeed data perform comprehensively better than all purely technical strategies.

Mean Reversion Strategy Doesn't Work in Perpetual Contracts

We tested three different versions of Agents based on the logic of “price deviates too much and will quickly revert” with the following performances:

• Viper: -18%
• Mamba: -33%
• Anaconda: -22%

The result was all losses. The issue lies in the fact that the trend in the Hyperliquid perpetual contract market is much stronger than the possibility of mean reversion. Attempting to catch a bottom in a downtrend is the costliest mistake in this market. These Agents constantly long at so-called “support levels,” but the price continues to decline for days.

We are testing a fix by adding a filtering tool for macro market conditions; that is, when the four-hour trend of BTC is a downtrend, it prohibits long positions based on the “mean reversion strategy.” Initial results look good; this filter can avoid 14 out of 28 losing trades for Mamba.

Don't Stick to a Single Model

Our latest Agent (Vixen) is based on Fox's trading data and adopts two entirely different entry modes.

• Stalker Mode: Capture the signal of smart money quietly accumulating a certain asset through multiple scans. This way, you can enter before the crowd rushes in; Fox's largest profits come from this mode.
• Striker Mode: Capture sharp breakout trends confirmed by trading volume. Enter as the market explodes, but only execute when real trading volume supports it (filtering out false pumps).

Fox's data suggests that these are actually two completely different sources of Alpha signals. If you only use a single entry mode, you must choose between the two, thereby missing out on the other opportunity.

Agents Self-Adjust — and the Result Is Always Worse

An unexpected finding is: When Agents experience consecutive losses, they try to “self-repair.” Common self-repair behaviors include loosening entry conditions, increasing leverage, and removing risk protection mechanisms, but the result is that each time losses accelerate.

For example, Dire Wolf activated 5 parallel leveraged positions at 25x after suffering a -27% loss and relaxed the order execution speed limit; another Agent removed the stagnation profit-taking mechanism; and another Agent raised its daily loss limit from 10% to 25%.

Our solution is to write risk protection mechanisms directly into the scanning tool code instead of relying on the Agents' own strategy configurations. If the scanning tool does not output signals, the Agent cannot execute trades — no matter how aggressive the adjustments might be in its own configuration.

Next Steps

We will continue running experiments for 24–48 hours and then shut down those Agents that no longer have a chance to recoup costs, in order to prevent remaining funds from continuing to deplete.

Next, we will deploy new versions of the strategy and write protection mechanisms into the code layer:

• Wolverine v1.1: HYPE Speed DSL trailing stop (faster profit locking in high volatility assets);
• Mamba v2.0: Mean reversion strategy + BTC macro trend protection;
• Scorpion v2.0: Real-time momentum event consensus (replacing outdated whale-following strategy).

At the same time, we will:

• Unify the strategy configurations of Fox, Vixen, and Mantis: These three Agents use the same scanning tool, but configurations have drifted; Fox currently has a yield of over 23%, while the other two will be adjusted to the same settings;
• Redeploy a new Fox/Vixen combination using Fox's complete winning configuration, including XYZ prohibition rules, stagnation profit-taking mechanism, 10% daily loss limit, and all risk gating mechanisms turned on;
• Expand the single asset hunter strategy: Grizzly's three-phase lifecycle model (Hunting → Harnessing → Stalking → Reloading) has now been applied to ETH (Polar), SOL (Kodiak), and HYPE (Wolverine).

Meanwhile, we are also developing brand new strategies and directly testing them in real markets. The market itself is the laboratory. Each new strategy will have $1,000 in funds and a fully transparent trading record.

Our experiments will run live at strategies.senpi.ai; all strategy codes are open-sourced at: github.com/Senpi-ai/senpi-skills

22 Agents, $22,000 in real funds, every trade fully transparent, the experiment continues.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。