How backtesting changes when you switch from manual trading to an agent

Manual backtesting is the process of reading historical charts and recording hypothetical entries and exits in a notebook or spreadsheet. An agentic backtest runs the same logic through code, applies exact position sizing in US dollars, and enforces hard budget caps and kill switches that paper trading alone cannot replicate. The difference is not just speed; it is whether your strategy survives contact with real execution, latency, and the emotional overrides that manual testing quietly permits.

What does manual backtesting actually look like?

When you backtest by hand, you scroll through historical candlesticks and ask yourself whether you would have bought or sold at each point. You write down the price, the date, and perhaps a brief reason in a column. You sum the profits and losses at the bottom to see if the idea is worth trying. Suppose you are looking at a daily chart for a stock. You see a moving average crossover, you mark the entry at the closing price, and you mark the exit five days later at the next cross. You repeat this for fifty signals and calculate a win rate. You might color the winning rows green and the losing rows red to build a visual sense of confidence. This method is accessible, requires no infrastructure, and gives you a tactile feel for how a market moves. It also creates an illusion that you understand the strategy deeply because you touched every data point. But that same intimacy lets you skip trades that look obvious in hindsight, change your position size retroactively, and ignore the fact that you might have been asleep or at work when the setup occurred. You can also pause the chart, look at the news for that day, and convince yourself you would have known to avoid the trade. Every manual backtest is a story you tell yourself about the past, and the narrator is usually sympathetic to the hero.

Where does manual backtesting usually break down?

The first failure mode is lookahead bias. You know what happened next, so you subconsciously refine your entry criteria until they fit the chart perfectly. A signal that looked ambiguous in real time becomes crystal clear when you already see the resulting trend. The second is execution fantasy. You assume you would have filled at the exact high or low of a candle, without slippage, delay, partial fills, or failed limit orders. You imagine your limit order resting on the book and filling cleanly, when in reality the price might have touched your level only briefly and moved away before your order arrived. The third is size inconsistency. A human tester might mentally risk two percent on one trade and five percent on another because the setup felt stronger, even though the written rules said otherwise. The fourth is omission. You forget to record the losing trades that happened while you were distracted, or you decide that a gap down was an anomaly that does not count. You might also exclude the first hour of the day because you are not a morning person, even though the rules did not specify a time filter. The fifth is emotional rewriting. After a week away from the spreadsheet, you look at a losing entry and convince yourself you would not have taken it because the volume looked wrong or the trend was weak. These errors are not cheating; they are the natural result of a human brain trying to hold state across hundreds of events. They are also the reason many strategies that look profitable in a spreadsheet lose money immediately when automated.

How does an agent change the test?

An agent does not feel sympathy. It reads historical data, but it applies the logic exactly as written, sizes every position in plain US dollars, and logs every intended trade whether it wins or loses. Because the agent has no memory drift, it cannot skip a losing trade because it was tired. Because it uses the API, it must account for the venue's actual minimum order sizes, tick sizes, and margin requirements. The API normalizes venue-specific contract math, so you size in dollars and the agent translates that into the correct lot size or contract multiplier. The test becomes an audit of the rules, not a memory of the market. If you tell the agent to enter when the short-term average crosses above the long-term average, it will do so on every cross, including the ones that immediately reverse. It will also apply the scoped key limits, budget caps, and position maximums you set before the test began. This means your backtest now includes the friction of your own safety infrastructure, which is something no spreadsheet can simulate. The agent experiences the data as a discrete sequence of events, not as a continuous scroll you can pause and reconsider. It must decide in the moment, with no ability to peek at the next bar. If your strategy relies on a vague sense of confirmation, the agent will expose that vagueness immediately because it cannot act on a feeling. It can only act on a condition. This rigor is uncomfortable, but it is the only way to know if the strategy is the rule set or the storyteller.

Why is paper trading with an agent different from a spreadsheet?

Paper trading through an agent is the bridge between historical simulation and live markets. In a spreadsheet, you assume you have unlimited buying power, no latency, and perfect fills. In agentic paper trading, the agent sends orders to a simulated venue that respects the same constraints as the live market. The agent must wait for a fill, manage a balance that goes up and down, and respect the daily or weekly budget cap you configured. Because the agent connects through MCP tools or the REST API, you can observe its reasoning in the same environment you use for other tasks, which lowers the barrier to supervision. You can watch the agent reason through its MCP tools, see the exact payload it would send, and review the audit log of every decision. This visibility means you are not just testing the strategy; you are testing the entire automation pipeline, from data ingestion to order formatting to error handling. This process reveals a class of errors that historical backtesting cannot catch. The agent might misinterpret a data feed, loop too aggressively between two prices, or request a size that exceeds your hard limit. It might attempt to trade a market that is closed or send an order denominated in a way the venue rejects. It might also misread a decimal place and request a size ten times larger than you intended. Catching these behaviors in paper mode costs you nothing but time. It also teaches you how the agent behaves when it is wrong. You want to see it hit a stop, sit idle, or flatten according to your plan, not override the rules because it is anxious. How to backtest AI trading strategies without fooling yourself covers the specific biases that paper trading should address before you add capital.

What safety checks should you add before the agent touches real money?

The transition from paper to live is not a single switch; it is a sequence of hardened limits. You should start with a scoped API key that can trade but can never withdraw funds. You should set a budget cap that, if breached, flattens positions and revokes the key automatically. You should define a position size limit in US dollars so the agent cannot concentrate your entire balance into a single options contract or a leveraged perps position. You should write an exit plan that tells the agent when to stop for the day, week, or month, regardless of whether it believes the next trade will be a winner. Finally, you should place a panic switch somewhere you can reach without opening a terminal. These controls are not obstacles; they are part of the strategy itself. They force you to decide your maximum pain threshold before the market decides it for you. Because Felix is non-custodial by construction, your funds sit in a wallet you control. The agent can spend within the limits you set, but it can never withdraw to itself or steal. Withdrawal addresses are owner approved only. You should also review the audit logs regularly to confirm that the agent is not drifting from its mandate, even within the hard limits. Live trading requires explicit owner authorization of a key, so you cannot accidentally connect a paper agent to real money without a deliberate step. This architecture means that even if your strategy is flawed, the damage is bounded by the caps you defined. How to start an AI trading agent with hard limits and How to run an AI trading agent with real money, safely describe how to implement each layer without trusting the agent to police itself.

How do you move from a backtest to live trading without skipping steps?

The honest path is to treat the manual spreadsheet as a hypothesis, the agentic historical backtest as a controlled experiment, and paper trading as a dress rehearsal with real constraints. Only after the agent has executed hundreds of paper trades without breaching limits, without lookahead errors, and without emotional overrides should you authorize a live key. Even then, the live key should carry a fraction of the budget you eventually intend to use. You increase capital only after the agent demonstrates that it can lose money gracefully, which means stopping when the daily cap is hit rather than doubling down to recover. Position sizing must be derived from the agent's actual buying power, not from the idealized balance in your notebook. You must also accept that the first live loss will feel different from a paper loss. The emotional weight of real money changes how you perceive the same drawdown, and if you have not hardened your rules in advance, you will be tempted to intervene manually at the worst possible moment. The goal is to make the live transition boring, which means the rules are so explicit that the agent needs no special permission to follow them. How to size positions for an AI trading agent from first principles explains why dollar-based sizing matters when an API normalizes across stocks, crypto, perps, options, and prediction markets. Remember that any backtest is a map of the past, not a promise of the future. Markets change, liquidity shifts, new fees appear, and any strategy can lose money, including the entire amount you allocated.