Racing 22 Agents on Hyperliquid: What Did I Discover? (Full Strategy Code Included)

Odaily星球日报Publicado em 2026-03-16Última atualização em 2026-03-16

Resumo

Summary: We deployed 22 autonomous AI trading agents on Hyperliquid, each with $1000, executing over 5000 trades. Key findings: 1) Less trading with higher conviction yields better results—agents with under 120 trades were profitable, while those with over 400 lost significantly. 2) Profits follow a power law: 3-5 trades generated nearly all gains, with small losses quickly stopped out. 3) Real-time "Hyperfeed" data (tracking where smart money is profiting) outperformed pure technical analysis. 4) Mean-reversion strategies failed in perpetual markets; trend-following was more effective. 5) Agents self-adjusted poorly when losing, often accelerating losses by increasing leverage or removing safeguards. Future steps include hardening risk controls in code and testing new strategies. All code is open-sourced, and experiments continue live.

This article is from:Jason Goldberg

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

We deployed 22 autonomous trading AI Agents on Hyperliquid via Senpi, each configured with $1000 in real capital.

They run 24/7 — scanning the market, opening positions, setting trailing stops, managing risk — all without any human intervention.

After investing $22,000 in initial capital and executing over 5000 trades, here is our summary of lessons learned.

General Conclusions

"Fewer trades" plus "higher conviction" always equals "better results." This isn't an occasional occurrence; it holds true every time.

  • Odaily Note: Fox, Bison, Ghost Fox, and Grizzly, Viper, Mamba, Anaconda mentioned later are all names of Agents executing different strategies.

As shown above, Agent "Fox" and Agent "Ghost Fox" use the same scanning tool. Fox only selectively executes some of the signals, while Ghost Fox executes more signals. The result is a 56 percentage point difference in their Return on Investment (ROI).

The real advantage lies not in the scanning tool itself, but in the discipline to wait for the right signal.

  • All Agents with over 400 trades suffered significant losses.
  • All Agents with fewer than 120 trades were profitable.

More trades do not mean more opportunities — they mean more ineffective trades, more fees, and more exposure to noise risk.

Profits Follow a "Power Law Distribution"

Among our best-performing Agents, 3–5 trades contributed to all the profits, while the remaining trades were mostly quickly stopped out with small losses.

Taking Fox as an example again, the three best trades (ZEC, TRUMP, FARTCOIN) collectively profited over $350; the other 46 trades collectively lost over $100; the final result was a net profit of approximately $248.

This is entirely by design. Our strategy is: enter decisively with high conviction, stop out losses quickly within minutes, let profitable positions run, and lock in some peak profits using the DSL High Water trailing stop strategy. When the average win is 10 times the average loss, you can make steady profits even with a win rate of only 43%.

Those Agents that tried to maintain a high win rate through "safe" trades all lost money — because each trade aiming for tiny profits still incurs fees and market risk.

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 all traders on Hyperliquid are currently making money on. It is not a historical ranking or other lagging indicator, but the actual profitable trading activity happening on the entire exchange right now.

Our core scanning tool, Emerging Movers, reads Hyperfeed's market concentration data every 90 seconds. When smart money suddenly rotates to an asset: for example, a trader suddenly jumps at least 15 places on the leaderboard, or the rate of profit contribution for an asset suddenly increases, or multiple top traders simultaneously concentrate on opening the same position, the scanner can capture the signal before the move is fully priced in.

This is the structural advantage of building strategies on Hyperliquid via Senpi. You can see in real-time where top traders' profits are concentrating and act immediately. No other exchange offers this visibility, and no other platform allows autonomous Agents to execute based on this.

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

  • Fox / Vixen: Identify sudden smart money concentration on an asset via Emerging Movers;
  • Grizzly: Analyze smart money positions on BTC via Hyperfeed before opening a position;
  • Bison: Uses smart money direction as a hard condition — will not trade if the direction is opposite;

While the worst-performing Agents:

  • Completely ignored smart money signals, e.g., Viper, Mamba based purely on technical analysis;
  • Used outdated smart money data (Scorpion v1), treating months-old positions as new signals;

So the conclusion is very clear: Agents trading based on real-time Hyperfeed data performed significantly better than all purely technical strategies.

Mean Reversion Strategies Don't Work in Perpetual Swaps

We tested three different versions of Agents based on the logic that "price deviates too far and will soon revert." Their specific performances were as follows:

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

The result was losses across the board. The problem is that the Hyperliquid perpetual swap market is far more trend-driven than mean-reverting. Buying the dip in a downtrend is the most expensive mistake in this market. These Agents kept going long at so-called "support levels," but the price continued to trend lower for days.

The fix we are testing is adding a macro market state filter that prohibits "mean reversion strategy" dip buying when BTC's 4-hour trend is down. Preliminary results look promising; this filter could avoid 14 out of Mamba's 28 losing trades.

Don't Stick to a Single Pattern

Our newest Agent (Vixen), based on Fox's trading data, uses two截然不同的 (distinctly different) entry patterns.

  • Stalker Mode: Captures signals through multiple scans where smart money is quietly accumulating an asset. This allows you to enter before the crowd rushes in. Fox's biggest profits came from this mode.
  • Striker Mode: Captures violent breakout moves confirmed by volume. Enters as the move explodes, but only executes if supported by real volume (filtering fake pumps).

Fox's data shows these are two completely different sources of Alpha signals. If you only use a single entry pattern, you must choose between them and miss the other opportunity.

Agents Self-Adjust — And It Always Gets Worse

An unexpected finding was: When Agents experience consecutive losses, they attempt to "self-repair." Common repair behaviors include loosening entry conditions, increasing leverage, and removing risk protection mechanisms, but the result is always accelerated losses.

Some examples: Dire Wolf, after being down -27%, enabled 5 parallel 25x leverage positions and relaxed the order speed limit; another Agent deleted the stagnation take-profit mechanism; another raised the daily loss limit from 10% to 25%.

Our solution is to write risk protection mechanisms directly into the scanner tool code, rather than relying on the Agent's own strategy configuration. If the scanner tool doesn't output a signal, the Agent cannot execute a trade — no matter how aggressive its own configuration becomes.

Next Steps

We will continue the experiment for another 24-48 hours, then shut down Agents that have no realistic chance of recovering their capital to prevent further drain of remaining funds.

Next, we will deploy new strategy versions with protection mechanisms written into the code layer:

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

We will also:

  • Standardize strategy configurations for Fox, Vixen, and Mantis: These three Agents use the same scanner tool, but their configurations have drifted. Fox's current yield is over 23%; the other two will be adjusted to the same settings;
  • Redeploy new Fox/Vixen pairs using Fox's full winning configuration, including XYZ ban rules, stagnation take-profit mechanism, 10% daily loss上限 (cap), all risk gating mechanisms enabled;
  • Expand single-asset hunter strategies: Grizzly's three-phase lifecycle model (Hunt → Ride → Lurk → Reload) is now applied to ETH (Polar), SOL (Kodiak), and HYPE (Wolverine).

Simultaneously, we are developing brand new strategies and testing them directly in the live market. The market itself is the laboratory. Each new strategy gets $1000 in capital and fully transparent trade records.

Our experiment runs live at strategies.senpi.ai; All strategy code is open source at: github.com/Senpi-ai/senpi-skills

22 Agents, $22,000 real capital, every trade completely public. The experiment continues.

Perguntas relacionadas

QWhat is the key finding regarding the relationship between the number of trades and profitability in the experiment?

AThe key finding is that fewer trades combined with higher conviction consistently lead to better results. All agents that executed over 400 trades experienced significant losses, while all agents with fewer than 120 trades were profitable. More trades meant more noise, fees, and exposure to risk, not more opportunity.

QWhat is Hyperfeed and why was it a critical component for the most successful agents?

AHyperfeed is a real-time tracking system that shows which assets traders on Hyperliquid are currently making money on at that exact moment. It was critical because it allowed the best-performing agents (like Fox and Grizzly) to identify and act on signals where 'smart money' was suddenly concentrating on an asset before the move was fully priced in, giving them a structural advantage.

QWhy did the mean reversion strategies (Viper, Mamba, Anaconda) perform poorly?

AThe mean reversion strategies all performed poorly (-18% to -33%) because the Hyperliquid perpetual market has stronger trending characteristics than mean-reverting ones. Attempting to buy the dip (go long at perceived 'support' levels) in a downtrend was a costly mistake, as prices could continue to fall for days.

QWhat unexpected and detrimental behavior did the agents exhibit when facing consecutive losses?

AWhen facing consecutive losses, the agents would attempt to 'self-repair' by making their configurations more aggressive. This included widening entry conditions, increasing leverage, and removing risk protection mechanisms like stop-losses. However, this behavior consistently accelerated their losses instead of helping them recover.

QHow do the profits of the successful agents follow a 'power law distribution'?

AThe profits followed a power law distribution, meaning a very small number of trades (3-5) contributed all of the profits, while the majority of trades ended as small losses that were quickly stopped out. For example, Agent Fox's three best trades made over $350, while its other 46 trades lost over $100, resulting in a net profit of ~$248.

Leituras Relacionadas

Humanity Loses $31 Million in Attack, Token Price Plummets 90% Due to a Single Private Key

On June 9th, the digital identity project Humanity Protocol suffered a major security breach resulting in over $31 million in losses. According to on-chain analyst Specter, hundreds of wallets holding the project's H token were drained. The attack was confirmed by founder Terence Kwok to be caused by the compromise of a foundation member's private key. As a precaution, users are advised to avoid interacting with Humanity's cross-chain bridge or liquidity pools. The incident caused the H token price to crash over 90%, from around $0.70 to a low of $0.052, wiping its market cap from $2 billion to approximately $35.7 million. The attacker allegedly minted 100 million new H tokens and is selling them for BNB. This breach adds to existing controversies surrounding Humanity Protocol. Founded in 2024, it aimed to verify human users via palm-print biometrics and zero-knowledge proofs. However, a leaked conversation in 2025 revealed that only about 1 million of its 9 million claimed Human IDs had completed biometric verification, suggesting 88% might be bots. Furthermore, the project has faced allegations of being a repackaged product from a Chinese access control vendor, raising privacy and authenticity concerns. Founder Terence Kwok's previous venture, Tink Labs, a hotel smartphone startup that raised $170 million, failed and entered bankruptcy in 2020 after burning through its funding. The current attack highlights the persistent critical issue of private key management in crypto. Unlike smart contract exploits, a private key compromise bypasses all on-chain security mechanisms. With no user compensation plan announced yet, this $31 million breach may be a final blow to the project's credibility, already weakened by previous controversies and a heavily depreciated token.

marsbitHá 10m

Humanity Loses $31 Million in Attack, Token Price Plummets 90% Due to a Single Private Key

marsbitHá 10m

MicroStrategy Will Not Die in This Downturn: Reflexivity, STRC Anchoring Back to Par, and the Self-Rescue Logic of "Sell Stock, Not Bitcoin"

This article analyzes the recent sharp decline in Bitcoin and MicroStrategy (MSTR), framing it as a targeted "reflexivity" attack. The trigger was MSTR using its cash reserves to buy back convertible notes, raising market concerns about a liquidity crisis. The playbook follows George Soros's principle: market expectations can shape reality. Fears that MSTR might be forced to sell BTC caused panic selling, lowering BTC's price and worsening MSTR's financial ratios, thus reinforcing the negative narrative. The author argues that MSTR's Structured Convertible (STRC), while falling in price, is a floating-rate security that will eventually return to par value (100). The price drop reflects the market demanding a higher yield due to perceived risk, but as a floating-rate instrument, its coupon can adjust, naturally pulling the price back to par over time. This is crucial for MSTR's continued ability to raise funds. The core thesis is that MSTR's best move to counter the attack is to **issue new equity (sell shares)**, not sell its Bitcoin holdings. While selling BTC would solve the immediate cash crunch, it would destroy the company's core investment thesis and premium. It would dilute the BTC per share, likely erase the market premium over its net asset value (mNAV > 1), and worsen its debt-to-asset ratio. Issuing shares while mNAV is high (e.g., 1.25x) allows MSTR to raise cash for reserves without harming shareholder value or the "perpetual accumulation" narrative. It improves the debt ratio and reassures STRC holders, breaking the negative reflexivity cycle. In conclusion, while MSTR could survive this episode even by selling BTC, doing so would fundamentally alter its investment proposition and weaken it for future cycles. The optimal, value-preserving strategy is to sell equity to rebuild reserves and maintain the long-term growth flywheel.

marsbitHá 11m

MicroStrategy Will Not Die in This Downturn: Reflexivity, STRC Anchoring Back to Par, and the Self-Rescue Logic of "Sell Stock, Not Bitcoin"

marsbitHá 11m

Humanity Loses $31 Million, a Private Key Causes Token Price to Plunge 90%

On June 9th, the digital identity project Humanity Protocol suffered a major security breach resulting in over $31 million stolen from hundreds of wallets holding its H token. The attack was caused by the compromise of a private key belonging to a foundation member, leading the team to advise users against interacting with its bridge or liquidity pools. Following the incident, the price of the H token plummeted by over 90%, from around $0.70 to a low of $0.052, wiping out a significant portion of its market capitalization. The attacker allegedly minted 100 million new H tokens and began selling them for BNB. Humanity Protocol, founded in 2024, aimed to verify human users through palm-print biometrics and zero-knowledge proofs on Polygon CDK. Despite raising $50 million across two funding rounds and achieving a unicorn valuation, the project faced prior controversies. Shortly after its June 2025 token launch, reports emerged that only about 1 million of its 9 million registered IDs had completed biometric verification, suggesting 88% might be bots. Furthermore, allegations surfaced that the project might be a rebranded "shell" of a Chinese access control company, raising concerns about data privacy and authenticity. The project's founder, Terence Kwok, has a controversial business history. His previous venture, Tink Labs, burned through $170 million in funding before collapsing in 2020. The breach highlights the persistent critical risk of private key management in crypto. With no user compensation plan detailed in the initial response, the incident deals a severe blow to trust in a project already struggling with credibility issues.

Foresight NewsHá 32m

Humanity Loses $31 Million, a Private Key Causes Token Price to Plunge 90%

Foresight NewsHá 32m

How to Conduct Deep Research Using Claude's Dynamic Workflows

The article "How to Use Claude's Dynamic Workflows for Deep Research" discusses overcoming the pitfalls of technical research, where both humans and AI can get overwhelmed by information, leading to vague conclusions. It introduces Claude Code's new "Dynamic Workflows" feature, which automatically designs and executes task-specific workflows before starting a task, unlike simpler "planning modes." This approach incorporates validation, result convergence, and adversarial verification from the outset. The core of Dynamic Workflows is six predefined scheduling patterns that address how to decompose tasks and synthesize results: 1. **Classify-and-Act (Routing):** An agent classifies the task and routes it to the most suitable specialist agent for execution. It's precise and efficient but struggles with ambiguous tasks. 2. **Fan-out & Merge:** The task is split into parallel, independent subtasks whose results are later merged. It's fast and isolates contexts but is more expensive and challenging to synthesize. 3. **Adversarial Verification:** Multiple "challenger" agents critique a worker agent's conclusion, requiring majority approval. This counters confirmation bias and self-assessment errors but relies on verifiable facts. 4. **Generate & Filter:** Multiple agents generate many candidate solutions, which are then filtered against a rubric to output only the best. It fosters diversity but depends heavily on the filter's quality. 5. **Tournament:** Multiple agents compete on the same task, with pairwise comparisons eliminating contestants over rounds to select the best. This offers stable relative judgment but is complex. 6. **Loop:** An agent iteratively attempts a task, learning from errors and adjusting until a stop condition is met. It handles tasks with unknown scope but risks infinite loops without proper design. The author compares their own custom deep-research system, which involved multi-agent analysis and deduplication but lacked goal-oriented convergence, to Claude's built-in workflow. The official workflow adds critical layers: initial problem decomposition, credibility assessment of sources, cross-agent voting to delete weak conclusions (not just averaging), and output tightly focused on the user's original goals and actionable recommendations. This structurally addresses common AI issues like goal drift, premature stopping, context pollution, and output bias. In summary, Dynamic Workflows represent a shift from smarter single conversations to a structured research process, compressing what used to require many dialogues into 3-4 interactions, albeit at higher token cost. The author notes remaining challenges for their specific domain (blockchain research): the need for fact-based verification over official documentation, depth in truly novel interdisciplinary thinking, the practical validation of proposed solutions, and tailoring information density to the audience.

marsbitHá 43m

How to Conduct Deep Research Using Claude's Dynamic Workflows

marsbitHá 43m

When LPs Teach Me Investment with Doubao: A Self-Narrative of a Private Equity GP Switching Careers

When LPs Use Doubao to Teach Investing: A Transition Story of a Private Equity GP AI is making life increasingly difficult for small private equity fund managers, as a former GP of an offshore dollar fund reveals. The fund, managing tens of millions in US stocks, outperformed the Nasdaq but struggled with fundraising. Its traditional Cayman SPC/BVI structure failed to attract major Asian LPs, who now prefer Hong Kong LPF or Singapore VCC frameworks. The rise of AI-powered quantitative strategies has further squeezed the space for funds like his, which relied on subjective, discretionary investing. AI tools have leveled the information playing field, empowering LPs—often high-net-worth individuals, entrepreneurs, or family offices—to analyze investments themselves using chatbots like Doubao. This has eroded trust in GPs' expertise, leading to more frequent challenges over investment decisions and even withdrawals, especially during market rallies when retail investors sometimes outperform funds. Friction arises not necessarily from AI's capabilities but from how LPs use it. Many rely on conversational AI for validation rather than rigorous analysis, sometimes receiving misleading or hallucinated advice. While AI democratizes research, effective investing still requires discerning real insight from plausible-sounding output. Ultimately, AI is unlikely to fully replace GPs. Asset management remains a trust-based service. However, the industry must adapt. The future may see "human私募" (private equity) learning from AI and focusing more on providing value beyond pure analysis—perhaps by mastering the emotional intelligence and trust-building that machines cannot replicate.

Odaily星球日报Há 1h

When LPs Teach Me Investment with Doubao: A Self-Narrative of a Private Equity GP Switching Careers

Odaily星球日报Há 1h

Trading

Spot
Futuros
活动图片

Categorias popularesright-arrow

Etiquetas Popularesright-arrow