Common mistakes with spend caps and drawdown limits for trading agents

Spend caps and drawdown limits are the most common safety controls people attach to trading agents, but they are frequently misdesigned because they treat a dynamic portfolio as a static bank account. A fixed dollar spend cap ignores the difference between cash deployed and economic exposure, while a percentage drawdown limit can be gamed by an agent that inflates peak value through illiquid marks before losing. When these limits are not scoped to individual strategies and markets, they fail to contain risk and can even hide it behind a single aggregate number.

What do spend caps and drawdown limits actually control?

Spend caps are usually set as a maximum dollar amount an agent can deploy into open positions over a given period. Drawdown limits are usually set as a percentage decline from a peak portfolio value that triggers a halt. Both are meant to be circuit breakers. They are easy to explain and easy to implement, which is why they are the first controls people reach for. But ease of implementation does not mean they map well to the mechanics of an agent that trades across stocks, crypto, perpetual futures, options, and prediction markets through a single API.

The problem starts with the definition of spent. In a single stock account, a spend cap might mean the total value of buy orders. In a cross margin perps account, the same notion of spend is entangled with leverage and collateral requirements. In options, a small cash outlay can control a large notional exposure. A naive dollar cap that simply sums order values will either be too loose in high leverage markets or too tight in cash markets. It does not measure economic exposure. It measures an administrative variable.

Drawdown limits suffer from a similar definitional blur. The portfolio value used to compute the peak is often a mix of last traded prices, mark prices, and model values. An agent trading across venues does not hold one portfolio in one currency. It holds a basket of exposures, some of which may not have reliable real time marks. A global drawdown limit therefore applies a single percentage to a number that is itself an approximation, and then uses that percentage to decide whether to halt a system that may be operating normally in one market while failing in another.

Why do fixed dollar caps fail across multiple markets?

When an agent trades through one API across five market types, a single dollar cap applied at the wallet level loses resolution. Suppose you set a ten thousand dollar cap. The agent could place five thousand dollars into a stock position and five thousand dollars into a perps position with ten times effective notional exposure. The cap is not violated, but the portfolio risk is far higher than the cap implies. The owner thinks the agent is limited to ten thousand dollars of risk, but the agent is actually managing fifty thousand dollars or more of economic exposure.

People often make the mistake of setting the cap based on the cash they are willing to lose, not based on the exposure the agent can generate. This is one of the common position sizing mistakes that owners make when letting an agent trade real money. The correct first principle is to size limits by the economic risk in the position, not by the cash required to open it. Felix normalizes order sizing into plain US dollars to reduce this confusion, but the owner still needs to translate that into exposure limits per market type.

Another failure mode is temporal. A daily spend cap of two thousand dollars sounds conservative. But if the agent trades in and out of a volatile prediction market twenty times in an hour, the total gross volume can be high while the net exposure flips sign repeatedly. The cap might trigger and halt the agent, or it might not, depending on whether you count gross or net spend. Most people do not specify which, and the resulting ambiguity lets risk leak through. If the cap is gross, the agent can be stopped by legitimate hedging. If the cap is net, the agent can build a massive offsetting position that looks small on a net basis but carries huge tail risk on one leg. The mistake is assuming that a single dollar number can capture the complexity of multi market flow without explicit rules for netting, hedging, and tenor.

How do percentage drawdown limits ignore the path of losses?

A drawdown limit is typically stated as a percentage decline from the highest watermark the portfolio has reached. The idea is that if the agent loses, say, twenty percent from its peak, something is wrong and it should stop. The first mistake is assuming the peak was real. An agent can manufacture a temporary peak through illiquid or mark to model positions. In a prediction market, a thinly traded contract might show an inflated valuation that sets a high watermark. The agent then trades normally, the mark falls, and the drawdown limit triggers even though no true economic gain ever existed. The owner is left with a halted agent and a confused understanding of what went wrong.

The second mistake is conflating time weighted drawdown with path dependency. Drawdown limits are usually checked at discrete intervals. An agent could lose ten percent, recover, lose ten percent again, and never trigger a twenty percent peak to trough limit, yet the compounded damage to the portfolio is severe. The limit sees the recovery and resets its anxiety, while the owner sees a steadily eroding capital base. A twenty percent drawdown limit does not mean the agent can only lose twenty percent of your money. It means the agent can only lose twenty percent from the highest point it ever reached. If that highest point was itself unstable, the limit is built on sand.

There is also the confusion between account drawdown and position drawdown. An agent might hold a stock that is down fifty percent but offset it with a perp that is up fifty percent. The account shows flat. The owner sees no drawdown and leaves the agent running. But the stock position is now a concentrated, deep underwater bet that the agent may refuse to cut because the account level limit is not breached. Scoped limits per position, per strategy, or per market type would catch this, but a global percentage does not. The account can be flat while the portfolio is fragile. The global drawdown limit gives a false sense of safety because it averages away concentration risk.

Why do unscoped limits break under multi agent systems?

If you run more than one agent, or one agent with multiple strategies, a global cap or drawdown limit becomes a commons problem. Agent A spends its budget and leaves nothing for Agent B. Or Agent B takes a loss that triggers the global drawdown kill switch, stopping Agent A while it is in a profitable position. This is the mistake of applying a single limit to a system that is not a single entity. This is why architecting a multi-agent trading system requires scoped keys and isolated budgets so that one agent cannot trigger a halt for another.

The first principle fix is to scope budgets and limits to each agent, each market type, and each strategy. The Felix API allows scoped keys so that one agent cannot spend another agent's budget. But owners often skip this step because it requires more upfront design. They set one wallet level limit and assume fairness will emerge. It will not. Agents do not coordinate unless the infrastructure enforces boundaries.

This is particularly dangerous in multi agent systems where one agent hedges another. Suppose Agent One is long stocks and Agent Two is short perps as a hedge. If the stock market drops, Agent One shows a drawdown. A global drawdown limit might halt Agent One, but Agent Two is profitable. The owner is left with an unhedged short perp position and a confused system. Scoped limits, per agent and per market, would allow the owner to set different thresholds for the hedging leg. A hedge is expected to lose while the primary position loses, so its drawdown limit should be calibrated to the correlation, not to an absolute account value.

The same logic applies to spend caps. If two agents share a single daily cap, an aggressive agent can exhaust the budget before a conservative agent has executed its planned trades. The conservative agent is then paralyzed, not because markets are unsafe, but because the budget was a commons. Scoping converts a shared pool into a set of isolated envelopes, which is the only way to run multiple strategies without silent interference.

How should you design limits from first principles?

Start with the axiom that an agent can lose all money it can access. Therefore, the only safe limit is one that limits the maximum economic loss the owner is willing to accept, measured in a unit that the agent cannot manipulate. That unit is not peak portfolio value. It is not gross order volume. It is the owner's base currency, US dollars, at risk in a way that cannot be hidden through internal transfers or mark to model accounting.

First, separate the budget from the exposure. The budget is how much cash the agent can move into positions. The exposure is the notional or delta equivalent of those positions. Cap both, but cap them separately. A budget cap of five thousand dollars and an exposure cap of ten thousand dollars means the agent cannot deploy more than five thousand in cash, nor hold positions that represent more than ten thousand in economic terms. This prevents the leverage problem where a small cash outlay controls a large position.

Second, use path aware limits, not just watermark limits. Instead of a single peak to trough drawdown, consider a rolling loss limit over a fixed window. For example, if the agent loses more than one thousand dollars in any rolling twenty four hour period, halt. This removes the incentive to game a peak and captures compounding small losses. It also makes the limit comprehensible in dollar terms rather than percentage terms.

Third, scope every limit to a specific key, market, and strategy. The budget for a prediction market agent should be separate from the budget for an options agent. The drawdown limit for a hedging strategy should be wider than the drawdown limit for a directional strategy, because hedging strategies are expected to show interim losses against their primary leg.

Fourth, always pair limits with a panic switch. A spend cap that takes hours to enforce is not a safety control. A drawdown limit that requires manual review is not a circuit breaker. The Felix infrastructure includes a kill switch that flattens positions and revokes keys. Limits should be seen as the warning track, and the kill switch as the wall.

How do audit logs and observability complete the picture?

Limits are only as good as the data that proves they are working. An owner might set a spend cap and drawdown limit, then never check whether the agent is bouncing off them repeatedly. If the agent is hitting its spend cap every day and stopping, that is a signal that the strategy is misaligned with the budget, not that the safety system is working well.

Observability means tracking not just the limit breaches, but the distance to the limit. If an agent is consistently within five percent of its drawdown limit, the limit is too tight or the strategy is too risky. If the agent never gets within fifty percent of its spend cap, the cap might be irrelevant. These patterns require logs that are scoped to the same keys and markets as the limits themselves.

You should also log the reason for each halt. A drawdown halt on a Tuesday should be distinguishable from a spend cap halt on a Wednesday. If the logs show alternating drawdown and spend cap breaches, the agent is likely cycling through risk and capital in a way that neither limit alone can catch. The owner needs to see the cadence of breaches, not just the fact of them. You can learn how to set this up in our guide on audit logs and observability for trading agents with hard limits.

Without this visibility, limits become talismans. They are set once and trusted forever. But markets change, strategies drift, and agents find new ways to interact with the boundaries you have drawn. Observability is the feedback loop that tells you when your first principles need to be revisited.

Frequently asked questions