What most people get wrong about prediction market trading with AI agents

Most people assume that an AI agent trading prediction markets only needs to forecast outcomes better than chance, but the real failure modes are poor position sizing, ignoring liquidity constraints, and treating binary contracts like continuous markets. From first principles, a prediction market is a wallet-controlled, non-custodial venue where the agent must manage finite capital, fixed expiry, and discrete payoffs under hard spending limits. The common mistakes are not about bad predictions alone. They are about mismatched incentives between the model's confidence and the owner's risk, compounded by order books that disappear near resolution. An agent that understands the event may still destroy value if it does not understand the instrument.

What makes prediction markets different from other venues?

A stock or a perpetual future can be held indefinitely in principle, and its price moves along a continuous range. A prediction market contract resolves to either zero or one when an event concludes, and the price is bounded between those two points. This changes the nature of risk. An agent that buys a contract at seventy cents is not betting on upward momentum. It is risking seventy cents to win thirty cents, plus the return of stake, if the event occurs. The maximum loss is the entire entry price, and the maximum gain is the remainder to one. This asymmetry means that standard momentum or trend-following logic, which works in continuous markets, is often meaningless here. The agent is not riding a trend toward an open-ended target. It is holding a position that will collapse to a terminal value on a known or estimated schedule. Many builders port strategies from perpetual futures or equities directly into prediction markets without adjusting for this terminal structure. The result is that the agent enters positions that would make sense if the instrument could overshoot one dollar or reverse gracefully, but instead the contract simply expires. There is no averaging down to a better cost basis if the price moves against the agent, because the price is capped and time is finite. Understanding this requires stepping back to agentic trading risk first principles. The agent's owner still controls the wallet non-custodially, yet the agent must reason about a payoff profile that is fundamentally different from a perpetual swap or a share of stock. Felix abstracts the venue mechanics, but the agent cannot abstract the mathematics of a binary outcome.

Why do agents confuse probability with edge?

A common design pattern is to feed the model a market price and a model probability, then instruct the agent to buy when the model is more optimistic than the market. This confuses probability with edge. If the model believes an event has a sixty percent chance and the market prices it at fifty percent, the naive expected value looks positive. But edge must be calculated after accounting for the bid-ask spread, trading fees, and the time value of money locked in the position until resolution. If the contract resolves in six months, the capital tied up in that ten-cent perceived edge could have been deployed elsewhere. Worse, the model's sixty percent is itself a distribution, not a guarantee. In a binary market, being right fifty-five percent of the time with equal payoffs is profitable, but prediction market payoffs are rarely equal. Buying at eighty cents requires an eighty percent hit rate just to break even before fees, and most agents do not receive explicit instructions to perform that arithmetic. The agent needs a prompt that forces it to convert probability into expected return, then compare that return against a hurdle rate and the owner's budget. Calibration is not enough. A model can be perfectly calibrated, meaning its sixty percent predictions happen sixty percent of the time, and still lose money if it systematically buys at eighty cents and sells at forty. The agent must be instructed to compare its forecast against the market price, not just against a naive fifty-fifty baseline. This is where prompt design differs from manual trading rules. A manual trader might intuitively sense that an eighty-cent entry is too rich, or might notice that the model has been overconfident in similar past events. An agent will not, unless the prompt explicitly encodes the break-even logic, a historical accuracy check, and the non-custodial spend limits that prevent it from chasing thin edge with oversized positions.

How does liquidity change the math for automated sizing?

Prediction markets, especially on niche events, often have thin order books. The mid-price may look attractive, but the book depth at that price could be a few hundred dollars. An agent that sizes orders in dollars without inspecting available depth will move the market against itself. A five-hundred-dollar order in a thin book might clear the entire bid side or lift the ask through several ticks, turning a theoretical edge into an immediate loss on entry. Felix normalizes order sizing in plain US dollars across venues, but the agent still needs to respect the liquidity landscape of the specific market. As the event approaches resolution, liquidity often dries up further. Participants exit, spreads widen, and the remaining flow is information-heavy. An agent that assumes it can flatten a position at will is making the same mistake as a trader who assumes a stop-loss order will execute at the desired price in a gap. A non-custodial architecture with spend caps prevents the agent from stealing funds, yet it does not prevent the agent from locking the owner's capital in an illiquid, expiring contract that cannot be exited. That risk is real, and the owner can lose money, including the full position size. In some cases, the agent may interpret a wide spread as an arbitrage opportunity rather than a warning about liquidity, and place orders that sit unfilled while the event resolves. The prompt should instruct the agent to measure the book depth at the intended entry and exit, and to reject trades where the depth is thinner than a fixed fraction of the position size.

Why do static prompts fail when markets approach resolution?

A prompt written for a perpetual futures agent often assumes continuous trading, averaging, and the ability to hold until some external signal changes. Prediction markets have a hard resolution boundary. As that boundary approaches, the price should converge to the outcome, but it often does so with violent dislocations. Static prompts that do not mention expiry, resolution time, or post-event settlement behave as if the market will exist tomorrow in the same form. An agent holding a position through resolution may discover that the contract has settled and the payout is already determined, or that trading has halted and the final price is frozen. If the agent was short a contract that resolved at one, the position is a total loss. If the prompt did not include an explicit instruction to flatten before a certain deadline, the agent will not flatten. This is not a bug in the model. It is a mismatch between the prompt's implied time horizon and the instrument's fixed life. Builders should treat resolution time as a hard limit, similar to a safety control, and build exit plans that trigger well before the event concludes. The distinction between halting new trades and flattening existing trades is critical. A prompt that says stop trading an hour before the election does not necessarily tell the agent to close positions. It may simply sit on exposure through the resolution. The safety model for MCP trading tools treats owner-defined boundaries as hard constraints, not suggestions. A resolution-aware exit plan is simply another boundary. The agent should no more ignore it than it should ignore a spend cap. Near resolution, model drift is also dangerous. A language model may read news headlines and shift its probability rapidly, but the market may already be closed or the spread may be too wide to act. The agent must be instructed to ignore late-breaking signals and adhere to the pre-set flattening schedule.

How should a non-custodial agent size positions in binary markets?

Because Felix is non-custodial by construction, the agent trades from a wallet the owner controls and cannot withdraw funds to itself. The owner sets budget caps, position limits, and a kill switch. Within those rails, the agent must still decide how many dollars to place on a binary contract. The correct approach is to treat each potential trade as an independent wager with a known maximum loss and a known maximum gain. Suppose the agent identifies a contract trading at forty cents where the model believes the true probability is sixty percent. The expected value is positive, but the stake is at risk of total loss forty percent of the time. The agent should size the position so that a full loss does not breach the owner's daily or weekly drawdown limit. This is a different calculation from sizing a stock trade, where the position can be resized, hedged, or held indefinitely. In a prediction market, the contract is either worth one or zero. The agent cannot average down in the same way because the price is bounded and the timeline is fixed. Sizing positions step by step means the agent must compute notional exposure, maximum loss, and the probability of that loss, then check the result against the owner's hard cap. If the proposed trade would consume too much of the remaining budget, the agent should pass, even if the edge appears large. Correlation is also a hidden risk. Multiple prediction markets on the same underlying event, or on highly correlated events, can appear diversified to the agent while actually concentrating risk. The owner may see five separate positions, but a single news event could resolve all of them against the agent simultaneously. Sizing must therefore account for portfolio-level binary risk, not just individual contract edge. A kill switch and panic flatten mechanism remain essential. If new information arrives that invalidates the model's probability, the agent may not be able to react fast enough in a thin market. The owner can then revoke access and flatten manually. This layered approach, budget caps plus instrument-aware sizing plus owner control, is what makes agentic prediction market trading feasible without custody risk.

Why is the cost of capital harder to see in prediction markets?

When an agent buys a stock, the cost of capital is implicit in the opportunity cost of holding that stock instead of cash. When an agent buys a prediction market contract, the cost of capital is explicit and locked. The funds sit in escrow until the event resolves, and they earn no interest and cannot be redeployed. This drag is invisible to most agents because the prompt does not include a hurdle rate. An expected value calculation that ignores the time value of money will overstate edge on long-dated contracts. A six-month contract with a five percent expected return may underperform a three-month treasury bill once the risk-free rate is considered. The agent should therefore be instructed to annualize or time-weight its expected returns, and to reject trades where the edge is smaller than the owner's cost of capital. Felix allows the agent to trade across stocks, crypto, perps, options, and prediction markets from a single wallet, so the opportunity cost is real. Capital locked in a prediction market is capital that cannot respond to a stock or crypto signal tomorrow. The agent must treat the owner's portfolio as a unified budget, not as isolated silos where each market is judged independently.