What beginners get wrong when they first automate options trades with AI agents

Automating options trades with an AI agent introduces failure modes that do not exist in manual trading. Most beginners assume the agent will handle Greeks, liquidity, and expiration the same way a human would, but an LLM reasons about text, not market structure. The result is often mispriced orders, unintended directional exposure, or positions that expire worthless because the agent lacked an explicit instruction to close. Understanding these mistakes before you authorize a live key is the difference between a controlled experiment and an unplanned loss.

Why do agents treat options contracts like shares?

Options are not leveraged shares. Each standard contract controls a fixed notional amount, usually one hundred shares of the underlying, and its value depends on time to expiration, implied volatility, and the underlying price. A beginner might prompt an agent to "buy one thousand dollars of calls" without specifying whether that means one thousand dollars of premium, one thousand dollars of notional exposure, or ten contracts. The Felix API normalizes order sizing into plain US dollars, but the agent still needs to know what a contract represents. If the prompt is vague, the agent may reason in terms of shares and produce a trade that carries far more delta or gamma than the user intended.

A related mistake is confusing premium with risk. A far out-of-the-money option might cost fifty dollars per contract, so a small budget buys many contracts. The agent sees the low price and assumes the position is safe, but the entire premium is at risk and theta decay accelerates near expiration. The user did not tell the agent to track Greeks, so the agent does not. It simply buys what the prompt described. If the underlying moves slightly in the expected direction but volatility collapses, the position can still lose money. An agent that is not explicitly instructed to reason about delta or to avoid holding through earnings is effectively trading blind.

The owner must state in the prompt whether the strategy is directional, volatility based, or income focused. You should also define the contract multiplier and the maximum delta equivalent you are willing to hold. For example, if you want the exposure of one hundred shares, you should tell the agent to target a delta near one, not to buy one contract. If you do not want the agent to hold positions through earnings, you must state that explicitly. An agent that scans news might see an earnings announcement as a catalyst and enter a trade precisely when implied volatility is highest. Without instructions to avoid new positions in the two days before an earnings report, the agent may buy premium that is crushed immediately after the event. The prompt must define not just what to buy, but what the option represents in terms of risk.

What happens when an agent ignores liquidity and wide spreads?

Options markets often have wider bid-ask spreads than their underlying stocks, especially for deep out-of-the-money contracts, far expirations, or names with lower volume. A human trader notices the spread and adjusts the limit price. An agent reading market data may see a last traded price or a midpoint and place a limit order there without understanding that the midpoint is not a tradable quote. On paper trading, the fill might simulate at that midpoint. In live trading, the order sits unfilled while the underlying moves, or it fills at a much worse price than the agent expected.

This is particularly dangerous for multi-leg strategies. An iron condor or vertical spread requires four or two legs to fill at prices that make sense as a package. If the agent places each leg independently because the prompt said "enter this spread," it may get filled on the short legs and miss the long legs, leaving the account with naked exposure. The user must instruct the agent to check spread width, to use marketable limits when necessary, and to avoid names where the option book is thin. The agent does not inherently know what "thin" means. You have to define it, for example by telling the agent to avoid contracts where the bid-ask spread is wider than ten percent of the mid price.

Another subtle error is confusing opening and closing orders. An agent trying to close a long call might accidentally sell to open a new naked call if the prompt does not specify "buy to close" or "sell to close." The API abstracts direction, but the agent still needs to know whether it is reducing or adding to a position. If your account is not approved for naked selling, the order may be rejected, or worse, it may be accepted if the system interprets it as a covered position against a different holding. You should require the agent to print the intended order type, the leg side, and the expected net delta before it submits. If the numbers do not match your risk model, the order should not go out.

Why is a missing exit plan more dangerous with options than with stocks?

A stock position can be held indefinitely. An option has a fixed expiration, and its time value decays nonlinearly. Beginners often port their stock exit logic to options, setting a stop loss based on the option's mark price or a percentage of the premium. This fails because an option can lose value while the underlying stays flat, simply through theta decay. A stop loss on the option price might trigger after the profitable edge has already decayed away. A ten percent stop on a one dollar option is only ten cents, but theta can erase twenty percent overnight on a Friday before expiration.

Worse, an agent with no explicit time-based exit rule may hold a long option to expiration. If the option is out of the money, it expires worthless. If it is in the money, the agent may not know whether to exercise, sell before close, or let it expire and accept assignment risk. Short options carry assignment risk at any time for American style contracts. An agent that is short a call and does not monitor ex-dividend dates or early assignment probability can wake up to an unexpected short stock position. A short put can be assigned into a long stock position that exceeds the intended delta budget.

Your exit plan must include both a time stop and a price stop. You should also tell the agent to close all short options before the underlying goes ex-dividend, or to avoid selling American options altogether if the strategy is not prepared for early assignment. Building these rules into the prompt is essential. You can use the framework in our practical checklist for exit plans to make sure nothing is implicit. The panic switch on your account can flatten everything, but if the agent is holding dozens of spreads, flattening may execute at poor prices during a volatile session. It is better to have the agent close positions methodically according to a written plan than to rely on an emergency halt.

How do budget caps break with multi-leg strategies?

Felix lets you set scoped keys and budget caps so an agent cannot spend more than a defined amount. This works well for single leg stock trades. It breaks easily with options spreads if the agent views each leg as an independent trade. Suppose you set a one thousand dollar daily cap and the agent tries to open a butterfly spread that costs three hundred dollars in net premium but requires buying two legs and selling two legs. The absolute notional of the legs might sum to two thousand dollars, or the agent might sequence the legs and hit a per-order cap on one of the long legs. The agent then holds a partial spread with broken risk geometry.

The fix is to size the strategy as a unit, not as individual legs. You should tell the agent to calculate the net debit or net credit of the entire spread and to verify that the net entry price fits within the budget. You should also understand how your drawdown limits interact with option mark prices. An option can swing from a small debit to a large mark value based on volatility, so a drawdown limit calculated on the option's current price may trigger at the wrong time. If you set a drawdown limit of ten percent on the account value, a short volatility strategy might show a temporary mark-to-market loss that is ten times the expected maximum risk at expiration. The agent may flatten the position and realize the loss before the edge plays out.

A credit spread illustrates this well. You receive a net credit, but the short leg is marked against you every time the underlying moves closer to the strike. The account statement may show a large negative position value even though the maximum risk is still the width of the spread minus the credit. If your drawdown logic treats the mark price as realized loss, the agent will exit a healthy trade. You must configure limits based on the strategy's defined risk, not on the mark-to-market of individual legs. Review the common failure modes in spend caps and drawdown limits before you set thresholds for options. Non-custodial controls prevent the agent from stealing funds, but they do not prevent the agent from spending your budget on poorly structured spreads.

Why does paper trading not prepare you for live options automation?

Paper trading is useful for testing prompt logic and API integration, but it does not simulate the friction of live options markets. Paper fills often assume a midpoint or last price. Live fills depend on the venue's market maker, the time of day, the underlying's volatility, and whether the order is marketable. A strategy that prints paper profits for a month may fail in live trading because the agent cannot get the same entry prices.

Paper trading also rarely simulates assignment. An agent that sells a call spread may never see the short leg assigned in paper mode, so the prompt never gets tested for that branch. When live trading begins, early assignment on the short leg can turn a defined risk spread into an undefined directional stock position while the long leg remains open. The agent may not know how to handle the resulting stock delta, and the account may now be exposed to overnight gaps.

Before going live, you should backtest the strategy logic against historical tick data and walk through edge cases like expiration, assignment, and gap opens. Our guide on how developers can backtest AI trading strategies explains how to test prompts against historical scenarios rather than just paper profit curves. Live trading also requires an explicit owner authorization of the key, which is a deliberate checkpoint to review these assumptions. Do not treat authorization as a formality. Use it to verify that the agent has handled assignment, spread width, and time stops in paper mode. If the agent has never closed a position before expiration in paper, it should not be authorized to do so with real money.

What should you check before letting an agent trade options with real money?

Start with a narrow scope. Authorize a scoped key that can only trade options, with a tight budget cap and a short expiration on the key itself. Write the prompt as if you are explaining options to an intern who has never traded. Before you authorize the key, verify the following:

• ·The prompt defines contract multiplier, delta direction, and maximum days to expiration.
• ·The agent knows whether it is allowed to sell premium or only buy.
• ·The exit plan includes a time stop and a rule for handling assignment.
• ·The kill switch flattens positions and revokes the scoped key.
• ·The agent has tested closing a spread in paper mode, not just opening it.

Test the kill switch. In a non-custodial system, the funds sit in your wallet and the agent can spend within limits but cannot withdraw. You should still verify that the panic switch flattens options positions correctly, including after-hours or pre-expiration exits. Options can go from liquid to frozen in seconds near expiration. If the agent is not allowed to trade during the last hour, say so in the prompt.

Review the risk controls in how to control the risks of non-custodial trading with real money to understand how revocation and flattening work when the underlying is moving fast. The API is one key and one interface, but the market is still options. The agent does not know what it is doing unless you state it precisely. A single missing sentence in the prompt can convert a defined risk spread into a naked position, or turn a small premium play into an unexpected stock holding. Options automation is manageable, but only if every assumption about contract value, liquidity, and time is written down explicitly.