Once you have taught it enough, your agent stops only watching and starts trading: on its own simulated account, with its own $100,000, by fixed rules, every 15 minutes. Its trades go through the same verified ledger as yours, so its record can be judged and checked the same way.
When it starts
Your agent starts trading after you have closed 30 trades that can teach it (with market context, and closed at a different price from entry), counted across all of your accounts. It needs you to be signed in, so its trades are on a ledger. Progress toward 30 shows on /agent.
What happens on each run
A scheduled job runs every 15 minutes. For each agent that has been taught enough, it:
- Retrains, once a day, on all of your closed teaching trades, with the same model as the agent panel. See Your agent.
- Settles its own account: resting triggers are replayed against price history, as for any account.
- Closes any of its positions held for more than 24 hours, at the market, marked Time limit.
- Decides in your three most-traded markets, as described below.
How it decides
In each of your three most-traded markets that are still listed on Hyperliquid:
- If it already holds a position there, it holds; one position per market at a time.
- Otherwise it computes the market's features from the last 90 one-minute candles and asks the model for its probability of up.
- At 60% or more it goes long; at 40% or less it goes short; in between it waits.
| Rule | Value |
|---|---|
| Confidence to act | 60% for long, 40% or less for short |
| Risk per trade | 1% of equity lost if the stop loss is hit |
| Stop distance | 1.5 times the expected one-hour move, between 0.4% and 5% |
| Take profit | Twice as far from entry as the stop |
| Position size | Sized to the risk, at most 1x equity |
| Leverage | 5x, or the market's limit if lower |
| Holding time | 24 hours at most |
| Markets | Your three most-traded |
| Skipped when | The position would be under $10, or its stop would sit beyond the liquidation price |
Sizing, step by step
expected one-hour move = vol15m x sqrt(60)
stop = clamp(1.5 x expected one-hour move, 0.4%, 5%)
size = min(equity x 1% / stop, equity)
stop loss = entry x (1 - stop) long (1 + stop for a short)
take profit = entry x (1 + 2 x stop) long (1 - 2 x stop for a short)Example: equity $100,000, BTC at $100,000, vol15m 0.0015, 64% confident
expected one-hour move = 0.0015 x 7.746 = 1.16%
stop = 1.5 x 1.16% = 1.74%
size = 1,000 / 0.01743 = $57,377 at 5x ($11,475 margin)
stop loss = $98,257 (loses about $1,000, plus fees)
take profit = $103,486 (makes about $2,000, less fees)Its notes
Every trade it opens carries a machine-written reason, for example: Agent: 64% confident long, model trained on 42 of your trades. Stop 1.74% away, target 3.49%. The ledger also stores its exact probability with the trade. In the open dataset its trades go in their own file, autonomous.jsonl, with these notes always included.
Watching it
The Trading on its own section of /agent shows:
- its status (learning, trading or paused) and teaching progress;
- when its model was last trained and on how many trades;
- when it started trading, and what it did, or why it didn't, on its last run;
- its open positions and closed trades, with its confidence, stop and target, and how each ended;
- its own readiness checks and closed-trade equity, judged on its trades alone.
Pausing it
Pause on /agent stops it from opening anything new. Open positions keep their stop loss and take profit and are still settled. Resume picks up on the next run.
The agent only ever trades its simulated account. It can't place real orders, and nothing it does is a recommendation.