How to start an AI agent with a small budget

Start with an amount you can afford to lose completely. Configure a scoped API key with hard budget caps, per position limits, and a kill switch that you control. Test the agent in paper trading mode, then authorize live trading only after you have reviewed every parameter and confirmed the safety controls. Let the agent trade across stocks, crypto, perpetuals, options, or prediction markets, but keep the budget fixed until you have observed its behavior under real conditions for a meaningful period. Trading can lose money, including everything, so the first allocation should be small enough that total loss is acceptable.

What counts as a small budget for an AI agent?

A small budget is an amount that, if lost entirely, does not affect your living expenses, debt obligations, or emergency reserves. For some traders this is a few hundred dollars. For others it may be a few thousand. The exact number is personal and should be set without reference to what others trade. What matters is the psychological and financial reality that you must be willing to see the balance go to zero without experiencing distress. AI agents do not have intuition. They can misinterpret a prompt, overfit to recent price action, or encounter a market regime that their instructions do not address. When you give an agent real money, you are paying for information about how it behaves when outcomes matter. Treat the first allocation as tuition, not as an investment that must return a profit. If you cannot afford to lose the full amount, it is too large. This constraint is more important than the strategy, the market, or the model behind the agent. A small budget forces you to focus on process and safety rather than returns.

How do you connect the agent without giving up custody?

Felix is non-custodial by construction. Your funds sit in a wallet that you control. The agent receives a scoped API key that lets it place orders and manage positions within the limits you define, but it can never withdraw funds to itself or steal. Withdrawal addresses are owner approved only. You can connect the agent through MCP tools from Claude, Cursor, and other MCP clients, or through the REST API directly. Trading over MCP from Claude, Cursor, and Codex walks through the connection options. The agent never holds your private keys. It cannot move funds off your account. It can only buy and sell within the scoped permissions you grant. This means that even if the agent's logic fails, its environment is compromised, or its underlying model generates a harmful instruction, the worst case is bounded by the budget cap and position limits you set in advance. Non-custodial design does not remove trading risk, but it removes the risk of custodial theft or exit scams by the platform. You remain the owner of the funds at all times. The API simply acts as a controlled bridge between your agent and the market.

Which safety controls should you configure first?

Before the agent touches live markets, set four layers of control. First, a total budget cap that limits the sum of all orders across every market type. This is the maximum amount the agent can deploy at any time. Second, a per position limit so no single trade consumes the entire budget. Third, an exit plan that tells the agent when to close a trade, either by a profit target, a loss threshold, or a time limit. Fourth, a panic or kill switch that flattens all positions and revokes the key instantly. How to build a kill switch your trading agent cannot override explains why this must be external to the agent and immutable from its perspective. Why AI agents need scoped API keys when trading real money covers the key model in detail. These controls are not optional extras. They are the foundation that makes a small budget experiment survivable. Without them, a single misinterpreted prompt can deploy the entire allocation into one position. The exact request schema is in the docs; the shape looks like this.

{
  "key_name": "agent-small-budget",
  "max_budget_usd": 500,
  "per_position_limit_usd": 50,
  "allowed_markets": ["stocks", "crypto", "perps", "options", "prediction_markets"],
  "withdrawal_whitelist": ["OWNER_ADDRESS"],
  "panic_switch": true
}

After configuring these controls, verify them in paper trading. Trigger the kill switch intentionally. Watch what happens. If the agent does not flatten or if the key remains active, do not proceed to live trading.

How do you size positions when capital is limited?

With a small budget, position sizing is the difference between surviving a losing streak and wiping out on the third trade. Felix normalizes venue specific contract math so orders are sized in plain US dollars. This removes the need to calculate lot sizes, contract multipliers, or decimal precision manually, but it does not remove risk. You still must decide what percentage of the total budget any single position can risk. Many beginners let the agent use too much capital too early, often because they want the account to grow quickly. This is a mistake. How to size positions for an AI trading agent from first principles describes the failure modes in detail. A sensible starting point is to risk no more than one or two percent of the small budget on any single trade. If the budget is five hundred dollars, that means a five to ten dollar risk per position. This sounds small, but it lets you survive ten to twenty consecutive losses, which is plausible when an agent is learning a new strategy or adapting to a market it has not seen before. Resist the urge to raise size after a win. The goal is to stay in the game long enough to evaluate whether the agent has any edge.

What is the right way to move from paper to live trading?

Felix offers paper trading for testing. Use it to verify that the agent interprets your prompts correctly, that orders flow through the API to the intended venues, and that the safety controls trigger exactly as you expect. Paper trading catches basic errors in logic, formatting, or prompt interpretation. It does not guarantee future results. Live markets involve slippage, partial fills, and liquidity gaps that paper environments approximate imperfectly. When you are ready for live trading, authorize the key explicitly. This is a deliberate action where you review the budget, the allowed markets, the per position limits, and the kill switch one final time. Do not skip this step or assume that paper success translates directly to live profit. Start with the smallest possible live size, even if it seems trivial. The goal of the first live trade is not profit. It is to confirm that the agent, the API, and the market interact exactly as you expect when real money is at stake. Observe the first few trades closely. Check that the order size in dollars matches your intent. Check that the position limit blocked an oversized order. Verify that the exit plan closed the trade automatically. Only after this confirmation should you let the agent run unattended for short periods. Even then, keep the initial unattended window brief.

Why should you avoid changing the strategy after every trade?

It is tempting to rewrite the agent's prompt or replace its strategy after a single loss. This is usually counterproductive. A small budget is meant to give you enough data to distinguish between a flawed strategy and normal statistical noise. If you change the rules after every trade, you are not testing the agent. You are testing your own emotional reaction to the last outcome. Set the strategy, the position limits, and the exit rules before the first trade. Then let the agent run for a set number of trades or a fixed time period before you make any changes. This discipline is easier to maintain when the budget is small enough that individual losses do not feel catastrophic. Document the reason for every change. If you cannot explain the change in writing, do not make it. Over time, this record will tell you whether you are improving the agent or merely fitting to recent noise. Patience with a small budget protects you from overfitting.

How do you review and adjust after the first week?

After the first week of live trading, audit every order the agent placed. Check whether it respected the budget cap, whether any positions were larger than intended, and whether the exit plans or kill switch fired correctly. Look for patterns in the market types that caused the most slippage or unexpected behavior. If the agent lost money, determine whether the loss came from normal market variance or from a bug in the agent's reasoning. For example, if the agent repeatedly doubled down on a losing position, your prompt or limits may need tightening. If the agent was profitable, do not assume this will continue. A small sample size over one week is not evidence of an edge. Markets can enter favorable regimes by chance, and a few winning trades may reflect luck rather than skill. Adjust the prompt, the position limits, or the exit rules before you even consider increasing the budget. Only consider a budget increase after a longer observation period, perhaps four to six weeks, and after you have documented the agent's behavior in different market conditions, including at least one drawdown. If the agent breaches any limit during this period, stop immediately and fix the controls before continuing. The review is not about judging profit. It is about verifying that the system behaved as designed.