What a 52% 5-Day HYPE Rally Taught Our Hyperliquid Agents - and the Fixes They Shipped

Over the last five days $HYPE ran from ~$40 to ~$62 - a 52% move, with the violent part compressed into the final 36 hours.

We run a fleet of autonomous AI trading agents at Senpi on Hyperliquid with real capital in live markets. A move like this is the most honest exam there is, so we ran the obvious audit: pull every agent's actual on-chain fills and ask, bluntly - who caught it, who didn't, and why?

And then we challenged the agents to improve.

One pattern explains almost all of it: the agents that made money held; the agents that scalped caught the right direction and made nothing. One agent bought within cents of the absolute bottom and sold 35 minutes later for a 0.4% scalp, then watched the next ~50% happen without it.

Backtests lie by omission. Live fire doesn't.

You can make almost any strategy look brilliant on historical data. What you can't simulate is the texture of a real move: the order that rests on the book and never fills, the daemon that quietly dies at 22:17 on a Tuesday, the stop that felt safe in a spreadsheet and amputates every winner at the knee. None of these appear in a backtest. All three appeared in HYPE - and each became a fix.

Fix 1 - Let winners run (DSL re-tuning)

Several agents locked a profit floor after the first ~1% price wiggle. That optimizes the typical trade and quietly caps the exceptional one - and in crypto the exceptional trade is the entire business. We widened the trailing-stop ladder on the worst offenders (Cheetah, Scorpion, Orca) to a "let it breathe" profile: no profit lock until a position is up meaningfully, with the trail only tightening as the move extends. You give back a little on the median trade to keep the rare, fat-tailed winner uncapped.

The live proof arrived within hours: on the same push to $62, the held positions stayed in and compounded, while a tight-stop agent took its +$16 and was flat again before the next leg. (More on that below.)

Fix 2 - Execution plumbing is strategy

One agent tried to long HYPE before the breakout. The order went out as a passive maker limit to save fees, price ran away, the order never filled - and the engine reported a failure while leaving a dangling order on the book. No position, no participation. On one agent alone we found a dozen of these - and every momentum agent running the same maker-only config was exposed to the same trap.

The lesson: a fill-optimization that's correct for a patient trade is catastrophic for a momentum trade. So we split the fleet by intent:

• Momentum / trend agents (Python, Condor, Bison, Dire, Jaguar) now guarantee their entry - if the passive order won't fill quickly, they cross the spread as a taker. A few basis points of fee is nothing against a missed trend.

• Mean-reversion / fee-arb agents (the contrarian faders) keep maker-only entries - their entire edge is the rebate, and skipping a trade is the correct outcome when liquidity isn't there.

We also flagged the underlying "leave the order resting" behavior to be fixed at the runtime level - cleanly cancel an unfilled order rather than let it fill later with no risk controls attached.

Fix 3 - "Running" is not "trading"

The most sobering finding: several agents were online - heartbeating, polling prices, dashboards green - but their signal scanners had died days earlier and never respawned. Two died within five seconds of each other, pointing at a platform event rather than bad luck. They sat out the rally not because they made a bad call, but because they were structurally incapable of making any call.

Observability that watches the wrong layer is worse than none, because it lies to you. The fixes:

• Restarted the dead scanners in a crash-survivable way so a shell or session exit can't silently kill them again (two are already back and ticking).

• Re-enabling scanner-level liveness checks so a dead brain pages us in minutes, not days.

• Filed the platform-level mass-restart event for root-cause with the infra team.

The same rally, two opposite right answers

The clearest proof this isn't one-size-fits-all: two of our agents traded HYPE at the same time, in opposite directions, and both were right.

Wolverine is a momentum agent. It missed the bottom, but re-entered the trend late - long at $58.65. On the old tight ladder it would have locked a floor after a ~1% move and been shaken out on the first dip.

After its self-improvement: on the new wide ladder, once it cleared +10% the stop ratcheted up to its exact breakeven entry and stopped tightening there - the position became risk-free (worst case: a scratch) with full room to keep running. It's currently +13% and still holding, trailing the trend.

Dog is a contrarian fader - its job is to bet against the crowd. With Smart Money 70% long an exhausted +52% parabola, it shorted the top at 10x. But a fader's edge is bounded: a snapback is a fast, limited move, not a trend. So its risk logic is the opposite of Wolverine's - tight and quick. It rode the dip to +10%, its trailing stop locked the profit, and when HYPE bounced it banked +3.8% in 16 minutes and stepped aside.

One held for the trend. One grabbed the snapback and got out. Same asset, same hour, opposite trades - and the correct exit logic for each was the opposite of the other's. A single "best" stop-loss would have been wrong for at least one of them. That's the entire case for tuning risk per strategy, not per fleet.

The point isn't that things broke. It's the loop.

Every one of these was diagnosed from live trade-and-audit data, root-caused, and fixed the same day, while the rally was still live - config changes merged and rolling out to the fleet, runtime issues filed with exact timestamps and reproduction cases. That's the compounding advantage of running real agents in real markets: each move the market makes is a free, brutally specific test, and each test makes the next version of the fleet harder to break.

Backtests tell you what might have worked. A live 52% rally tells you exactly where you bleed - the dead scanner, the unfilled order, the stop that strangles your winners. We'll take the second one every time.

Get in the game: deploy your Hyperliquid Agent

Deploy your own agent at senpi.ai.

Track our fleet of open source strategies trading with real capital at strategies.senpi.ai.

Join the $100,000 trading competition at senpi.ai/arena.