Thesis
The most reliable information in a market is not a forecast. It is what the largest, most consistently profitable traders actually hold.
HyperSentinel reads those positions from the exchange, checks whether enough of them agree, and acts only when the agreement is clear and the market is moving. When neither condition holds, capital waits, or earns funding in a market-neutral pair.
01What it is
A non-custodial trading agent for Hyperliquid. You approve a trade-only key. The agent opens and closes positions on your own account alongside a curated set of large, proven traders. Funds never leave your account, and the key cannot withdraw. The agent is free today.
The moving parts
- Top traders. Around 8,600 screened addresses across roughly 280 markets.
- Positions, not trade tapes. Live account state on the exchange is the ground truth, alongside trend quality, funding, order books, marks and hourly bars.
- Macro regime. A bull / flat / bear read of the wider market, with a panic-rebound rule still in research.
- Hyperliquid. Orders, fills and funding happen on the exchange. USDC stays in your name. The agent key can only trade. Withdrawals need your wallet.
- The agent. Watches the data 24/7 and runs five checks in order: consensus (side, breadth, time), market gate (trending, not chop), no chasing, size and leverage with a protective stop, and idle capital into funding-rate positions.
How it works
02Where the signal comes from
- A screened universe of approximately 8,600 addresses across crypto, commodity and equity perpetuals, roughly 280 instruments.
- Positions are read from the exchange as they are, not rebuilt from trade tapes. Exchange account state is the truth.
- Each instrument gets a support share: how much screened capital leans one way, how many independent owners agree, and how long that agreement has held.
03Consensus
An instrument qualifies only when screened capital leans clearly to one side, the lean is shared by several independent owners rather than one large book, and it has held long enough to be a decision rather than a flicker.
When that agreement breaks, because owners reduce, flip or leave, the position is closed. There is no target price and no indicator. The exit is the same fact in reverse.
04Market gate
Consensus alone is not enough: a crowded position in a sideways market is a fee sink. A trend-quality gate keeps the agent out of chop. It asks whether the market is actually moving, and whether it got there in a straight line rather than as noise. In a flat regime no entries are made, whatever the consensus says.
05Macro market
Above the trend gate sits a read of the broader market, used in two ways.
Direction
The macro state is classified as bull, flat or bear. Longs open only in bull, shorts only in bear. A flat state gets no new entries.
Price of admission
Even in a confirmed trend the agent does not chase. If price has already run far from its recent range, the entry is skipped. In practice this is the gate that most often keeps the agent out: a rally everyone can see is usually already too far to buy.
Panic rebounds (research)
From a decade of daily macro data, one condition survived out-of-sample: a spike in market fear together with a sharp multi-day drop. It marks a rebound rather than a trend reversal. It is being evaluated as a separate, longer-holding lane and is not part of the live agent.
06Execution and risk
- Size follows the source's conviction, scaled to your account. Leverage per instrument mirrors the source.
- No chasing. An entry is skipped if price has already moved away from where the source built the position.
- A protective stop is placed on the exchange for every position, so it holds even if the agent is offline.
- Every order is written down as an intent before it is sent. Fills are reconciled against the exchange; any mismatch pauses new entries while protection stays on.
07Idle capital: earning funding rates
While the gate is closed, idle USDC can be placed market-neutral: long the spot asset, short the same perpetual. Price moves cancel; the short earns the funding rate every hour.
Candidates must show a stable positive funding history, limited time below zero, and deep enough books on both legs. The pair unwinds the moment a consensus entry is due, when funding stays negative, or when the market turns to a trend. Entry and exit cost trading fees, which set the minimum sensible holding time.
08The telemetry ledger
Every hour the agent writes down what it saw and what it did: the positions of the screened traders, the state of the market, funding, its own intents, fills and exits, each with the reason. This compounds as an asset. Live decisions are made against it, and every rule is re-derived from it rather than tuned by hand.
- Collect. Positions, fills and funding of the screened universe, hourly market state, and every order the agent placed, with the reason.
- Structure. Each instrument-hour becomes one record: who holds it, how much capital leans which way, how long the lean persisted, what the market and macro regime were, and whether anything looked unusual.
- Replay. Any candidate rule is run through the full history exactly as it would run live, then judged against a random-entry benchmark over the same period, on split halves, with confidence intervals.
- Promote. Only rules that survive all of that reach the live agent. Live outcomes then feed the ledger again.
Sample record
What one record looks like: an instrument-hour from the ledger. Owner identities are hashed and never leave the collector.
instrument_hour
venue HL · main
instrument BTC-PERP
observed_at 2026-09-18T17:00Z
consensus
side long
support_share 0.853 # share of screened capital on the winning side
owners 6 # independent books behind it
specialists 3 # owners with a track record in this instrument
persisted_for 27m # how long the lean has held unbroken
vs_peak 0.48 # support against its own high
regime
macro bull
trend_quality 0.52
extension +8.0% # distance from the recent range (chase filter)
anomaly_flags [early_specialist, distribution_watch]
decision wait # gate open · consensus ok · extension too far
outcome_24h — # filled in later; this is what replays learn from
our_order
intent_id 7f3a… # written before the order is sent
instrument xyz:GOLD
side long · unit 1 of 2
reason consensus_persisted · regime_ok · extension_ok
fill 4402.4 · 0.0029 · taker
exit consensus_decay · +0.55% · 142h
reconciled exchange state == ledger ✓
09Research and development
The agent is strict because the research said so. Most of what was tested did not survive. The rules that run live are what was left, not a first draft.
Why the gate is strict
On Hyperliquid's hourly BTC history, the trend gate has been open only about 12.5% of hours over the last 200 days (6.5% bull, 6% bear) and shut the other 87.5%. In the validation window the gate turned a losing sideways regime into a flat one, raised the profit factor from below two to above three, and cut the number of trades roughly in half.
How rules are chosen
Seventy-six gate configurations were replayed over four months of live consensus data and ranked by how well they transfer, not by the best single cell: both halves positive, still positive with the three best trades removed, bootstrap interval clear of zero, and ahead of a random-entry benchmark. The best-looking cell failed the split-half test and was rejected.
Copying timing does not transfer
Copying the exact entry timing of the screened traders was tested and rejected. They average down, and their timing lives inside their own risk budget; copying it turned a winner into a loser. What transfers is the fact of their positioning. Consensus selects the instrument; the market decides the timing.
Macro
Ten macro conditions were tested on ten years of daily data across eighty instruments. One survived: a spike in fear together with a sharp multi-day drop. It fires about once a year and marks a rebound, not a reversal. It is being built as a separate lane.
Regimes, not averages
Every result is reported by market regime (pullback, sideways, rally), because the winners of a crash are usually permanent bears and a rally's heroes are usually leverage. A rule that only works in one regime is labelled as such and never averaged into a headline number.
Exits and funding positions
Exit rules, stop widths, no-chase thresholds, funding-position filters and holding times were each gridded and replayed the same way. The live values sit on plateaus, not on peaks.
10Trust model
- Custody. Your Hyperliquid account. The agent holds a key with trade-only permission. Withdrawals require your own wallet signature.
- Revocation. You can revoke the key on the exchange at any time. The agent detects it and stops.
- Money in and out. Funding and withdrawal move through bridge solvers in seconds. Every step is signed by you, in your wallet.
- Transparency. The site publishes live aggregates of accounts running the agent, and the assistant answers from recorded state, never from a model's guess.
11Economics
The agent is free today. The planned model is a 15% success fee on realized profit above your account's high-water mark. Nothing in flat or losing periods, no management fee, no fee on principal. Collected fees fund the buyback of the HyperSentinel token, so the operator earns only when users do, and the proceeds return to the product's own asset. Terms and dates will be published before any fee takes effect.
12Sentinel Flywheel
The HyperSentinel token launches on pons. Each trade pays a 1% fee. The creator's 0.7% does not go to a wallet: it goes into the vault the agent trades and becomes NAV. A share of the agent's realized profit buys the token back on the open market and burns it, and that share ramps up with the size of the vault. Volume feeds the vault; the vault burns the supply. Nothing is distributed to holders.
The buyback share is a function of vault size. A small vault keeps its profit compounding; a large vault burns aggressively.
bmax is the ceiling. N0 is the vault size where the share reaches 63% of it. Both are fixed at launch and published here.
Illustrative model
Every input is a slider, none is a measurement. The front page shows the vault NAV live; buyback and burn counters start with the token launch.
13What it does not do
- It does not predict prices and does not trade on indicators.
- It does not run on accounts with other software or manual positions.
- It does not promise returns. The live figures on the site are what the agent actually did, including the months it waited.
14Status
Live on Hyperliquid mainnet with a small number of accounts. Funding and withdrawal paths, funding-rate positions and protection are exercised daily. Next: broader onboarding, the fee mechanism, and the token.
Working document. Figures are as of the date above and change with the market. The live site is the source of truth.