Risk limits

Risk limits are the rules you set when you create an agent. They are enforced by the engine on every order, whatever the model says and whatever the instructions say. The AI proposes; the limits decide what is allowed to happen.

How limits are applied#

• Each action is checked on its own, then the portfolio as it would look afterwards is checked. An action that would break a portfolio limit is rejected and not applied; later actions only see the approved ones.
• Risk-increasing orders face every limit. That means opening, increasing and non-reduce-only limit orders.
• Reducing orders face almost none. Closing, reducing and reduce-only orders only need the market to be open with a fresh price, so an agent can always get out.
• Every check is recorded on the decision, pass or fail, so you can see which rule blocked what.

The limits#

Max leverage#

The highest leverage the agent may use. It applies to each position and to the whole account (total position size divided by account equity). It can never exceed the market's own maximum.

Example: max leverage 3x on a $10,000 account. The agent may hold up to $30,000 of positions in total. An order asking for 5x is rejected with a message like "5x > max 3x".

Max position size#

The largest notional (position value in dollars) the agent may hold in a single market, counting what it already holds. In the wizard you set it as a share of the paper account.

Example: max position $2,500. The agent holds $2,000 in NVDA and asks to add $1,000. The result would be $3,000, so the add is rejected. Adding $500 is fine.

Max open positions#

How many markets the agent may hold at once. Example: max 3. With positions in NVDA, SOL and SPY, an order to open BTC is rejected; adding to NVDA is fine.

Free collateral#

Not a setting, but always checked: an order needs enough free margin (with a 1% safety margin). If the account does not have it, the order is rejected as insufficient collateral.

Allowed markets#

The list you ticked in the wizard. Anything else is rejected. An empty list means every enabled market. Example: you allow only SOL-PERP and ETH-PERP; an attempt to open NVDA fails with "not on the owner's market list".

The automatic brake#

Two limits act as a brake. When one trips, the agent is marked Stopped by a risk brake: all new risk is blocked, but exits keep running (stop-losses, take-profits, closing orders). You are notified once per day when a brake turns on.

Daily loss limit#

Measured as profit and loss since 00:00 UTC, as a percentage of the equity the agent started the day with. Deposits and withdrawals are excluded, so adding money does not hide a loss.

Example: limit 5%, the agent started the day at $10,000. At -$500 (-5%) the brake turns on. The day's clock resets at 00:00 UTC, and when the measured loss is back inside the limit the brake lifts by itself.

Max drawdown#

How far the agent's value may fall from its recent peak. The peak is looked for over the last 30 days and is deposit-neutral (money you add is not counted as gain).

Example: limit 20%, peak $12,000. At $9,600 the brake turns on and new positions are blocked. It lifts again when the drawdown shrinks below 20%, either because the agent recovers (for example exits at a profit) or because the old peak ages out of the 30-day window.

Both limits can be switched off in the engine (a limit left empty is off), but the wizard always sets them. A brake needs the agent to be active; it never overrides a paused agent.

Exit limits#

Widest stop loss#

The largest allowed distance between entry and stop, in percent. It applies to stop-losses and to trailing-stop distances. Example: widest stop 8%. A stop 12% below entry is rejected with "Stop 12.00% away > max 8%"; a 6% stop passes.

Reject orders without a stop loss#

When on, an order that opens or adds risk must come with a stop (a stop-loss or a trailing stop), or the position must already have one. Example: with it on, "open SOL long" with no stop is rejected. The keeper's own exit orders are exempt.

Let the agent move its exits#

When off, the agent cannot change stops, take-profits or trailing stops after the fact. When on, it may, but only tighten what you set: if you put the stop at 5% below entry, the agent may move it to 3%, never to 8%. An owner-set stop always wins. See exits.

The event window rule#

An optional limit: no new positions within a set number of minutes (1 to 1,440) before a scheduled event that concerns the market. Earnings dates apply to that stock; macro releases such as CPI, FOMC and the jobs report apply to every market. Exits are never blocked.

Example: window 60 minutes. NVDA reports in 40 minutes: opening or adding to NVDA is rejected with the event named; closing is allowed. Projected (unconfirmed) dates count too and are marked as projected. This rule is not part of the creation wizard; it is part of the risk settings the engine understands and can be supplied when an agent is created through the API (see bring your own agent).

Real-money pilot caps#

On top of your own limits, real-money agents run under platform caps during the pilot. They exist so that a mistake stays small.

These are pilot defaults and may change; the app shows the values in force. Reduce-only exits (stops, take-profits, trailing, your own closes) are never blocked by the caps or the kill switch, and you can always withdraw. See going live.

Choosing sensible limits#

• Start at 1x to 2x leverage and a small position size. You can loosen later; a loss cannot be undone.
• Set the daily loss limit to an amount you can shrug off, and the drawdown limit to the point where you would want to look at the agent anyway.
• Keep "reject orders without a stop loss" on, with a widest stop you could live with.
• Allow few, liquid markets until you have seen how the agent behaves.