Common position sizing mistakes when letting an AI agent trade real money

AI agents executing trades with real money often misjudge position size because owners confuse notional exposure with risk exposure, forget to translate backtest allocations into live capital limits, or allow the agent to infer size from wallet balance rather than a fixed budget. These errors compound quickly across stocks, crypto, perps, options, and prediction markets. Trading can lose money, including everything, and position sizing errors are one of the fastest ways to do so. The result is that a single signal can commit far more capital than the owner intended, or concentrate risk in one market without a mechanism to reduce it.

Why do agents trade larger than their owners expect?

The most common mistake is instructing the agent to use a percentage of available capital without defining what that capital means. An owner might say, "use ten percent of the wallet per trade." The agent reads the wallet balance, calculates ten percent, and submits an order. If the wallet holds ten thousand dollars, the first trade is one thousand dollars. If that trade wins and the wallet grows to eleven thousand, the next trade is eleven hundred. The owner did not intend to increase risk, but the sizing rule mechanically escalates it. This is especially dangerous in crypto and perps, where wallet balances can swing from collateral or unrealized profit before the agent places a new order. The agent sees a larger number and scales up.

Another source of drift is the difference between account balance and buying power. A stock broker might offer margin, and a perps venue might allow isolated or cross margin. The agent may see total equity rather than free cash, or may misread settled versus unsettled funds. If the owner does not specify which balance to use, the agent defaults to the largest available number. The fix is to give the agent a fixed dollar budget that is independent of the wallet balance. For example, "you may commit up to five hundred dollars per signal, regardless of the wallet size." This removes ambiguity and aligns with the safety practices for running an AI agent with real money.

What is wrong with dynamic percentage sizing without a ceiling?

Some owners believe that percentage-based sizing is safer because it automatically reduces exposure during drawdowns. While it is true that a ten percent rule will shrink the dollar amount as losses accrue, the same rule expands the dollar amount during winning streaks. After several profitable trades, the agent may be betting sums that would have seemed reckless at the start. The owner never updated their risk tolerance, but the sizing algorithm did.

There is also a psychological gap. A human trader might feel discomfort as position size grows and manually intervene. An agent does not feel discomfort. It follows the prompt. Unless the prompt contains a hard ceiling, such as "never exceed two thousand dollars on any single position," the agent will keep scaling. The absence of a cap means the agent can accidentally pyramid into a concentrated bet. A single bad move in an options venue or a perps venue can erase prior gains and more. The correct approach is to pair any percentage rule with an absolute maximum. For instance, "size at ten percent of the wallet, but never more than one thousand dollars and never less than one hundred." This bounds the behavior.

Why does leverage amplify sizing errors?

Felix normalizes order sizing in plain US dollars, which helps owners think in terms they understand. However, the underlying instruments still carry their own margin rules. When an agent sends an order for one thousand dollars to a perps venue, that one thousand dollars may represent margin for a notional position worth many times more. The owner thinks they risked one thousand dollars. In reality, they risked liquidation on a five thousand or ten thousand dollar exposure, depending on the leverage setting.

Options venues create similar confusion. An agent buying calls for one thousand dollars controls notional exposure tied to the underlying price and the delta of the contract. The owner sees the one thousand dollar debit and assumes the risk is capped at that amount. While the premium loss is capped, the notional exposure can still distort portfolio balance if the agent also holds offsetting positions in stocks or prediction markets. The agent does not automatically hedge across market types unless the prompt explicitly requires it.

The mistake is treating the order value as the risk value. Owners should size positions based on the notional exposure they are willing to bear, then translate that into the margin or premium required. If the owner wants no more than two thousand dollars of notional exposure to a crypto asset, and the perps venue offers ten times leverage, the agent should send an order sized to two hundred dollars of margin, not two thousand. The prompt must make this distinction clear, or the agent will overshoot. Building guardrails for the trading agent should include explicit leverage and notional limits, not just order limits.

How should a fixed dollar budget change agent behavior?

A fixed dollar cap is the simplest way to make an agent predictable. Instead of asking the agent to think about percentages or wallet balances, the owner gives it a budget. "You have one thousand dollars per trade and five thousand dollars total exposure." The agent can then track its own committed capital against that budget. When the budget is exhausted, the agent stops opening new positions. This prevents the common failure mode where an agent keeps opening new trades because it sees unused buying power.

Felix supports this through scoped keys and budget caps. The owner can create a key that is only allowed to commit up to a certain amount. Even if the agent's prompt is poorly written and asks for a larger size, the infrastructure rejects the order. This is a backstop, not a replacement for a good prompt. The owner should set the scoped limit slightly below the intended limit to allow for fees or slippage. If the agent is supposed to stay under one thousand dollars per trade, the scoped key might be set at nine hundred and fifty dollars to enforce a tighter boundary.

Fixed sizing also makes backtesting and paper trading comparable. When an owner tests a strategy in paper mode, they can see exactly how a one thousand dollar position behaves. If they later switch to live trading and change the size to a percentage of wallet, the paper results are no longer relevant. Consistency between paper and live sizing is critical. The article on how backtesting changes when you switch from manual trading to an agent covers this in more detail, but the core idea is that an agent needs invariant rules to produce invariant results.

What do exit plans and kill switches have to do with position size?

Sizing is an entry decision, but its consequences are felt at exit. An oversized position is only a problem if it moves against the owner. The problem is that the owner cannot predict direction, so they must assume the worst. A position that is too large will force a decision under stress. The agent may have a stop loss, but in volatile markets, especially in crypto or perps, slippage can turn a planned two percent loss into a much larger one.

Every position should have an exit plan before the entry is sent. The plan can be a stop, a time limit, or a take-profit level. The agent should be instructed to flatten if the plan is violated. More importantly, the owner should have a kill switch that flattens all positions and revokes the agent's access. This is not a sizing tool, but it is the final defense against a sizing error that has already occurred. If the agent opens a position that is ten times too large, the owner needs a way to stop it without negotiating with the agent.

Position limits also help. The owner can configure the agent to hold no more than three positions at once, or to avoid any single market above a certain percentage of the total budget. This prevents concentration even if each trade is within its dollar cap. The non-custodial trading checklist includes a step-by-step review of these controls, and it is worth completing before any live key is authorized.

How can paper trading reveal sizing mistakes before live capital is at risk?

Felix offers paper trading so that owners can observe an agent's behavior without committing real money. The mistake many make is to run paper trading with loose or unlimited sizing, then impose limits only when going live. This defeats the purpose. The paper environment should mirror the live environment exactly, including the fixed dollar caps, the maximum number of positions, and the scoped key behavior.

Watch the agent for a full week or through a complete set of market conditions. Does it respect the five hundred dollar cap when it sees a strong signal? Does it try to split one large order into two smaller ones to circumvent a per-trade limit? Does it correctly handle the transition from prediction markets to perps without doubling the intended exposure? These behaviors are easier to fix in paper mode.

Only after the agent has demonstrated disciplined sizing for an extended period should the owner authorize a live key. Even then, the live key should start with a smaller budget than the paper key used. Scaling up after proof of discipline is safer than scaling down after a loss.