Learn
AI trading agent without live execution
An AI trading agent can be useful before live execution is ever involved. Keep the workflow simulated, version the prompts, record paper decisions, and review risk behavior before treating any output as evidence.
Simulated agent workflow only
Trading Boy does not execute live trades, hold funds, or provide financial advice. This page is about evaluating AI-assisted paper decisions, not building a live trading bot or signal service.
What the agent should do in paper mode
A paper-mode AI trading agent should help organize context, apply written rules, produce reviewable rationale, and capture decisions in a journal. It should not be rewarded for sounding confident.
The first job is to make the decision process explicit. The prompt should say what markets the agent may review, what setup qualifies, what invalidates the idea, how paper risk is capped, and what fields must appear in the journal. The agent should be judged on whether it follows those instructions, not on whether a single paper trade made simulated profit.
The second job is to create comparable samples. Every test should record prompt version, persona version, market frame, rule version, and review window. Without versioning, the team cannot tell whether a better paper result came from an improved rule, a different market regime, a lucky outlier, or a hidden behavior change.
The third job is to keep live execution out of scope. Even when the paper result is strong, the review should continue to mention that paper trading cannot prove live fills, slippage, liquidity, fees, emotional response, or future returns.
Evaluation checklist
| Evaluation area | Paper-mode question | Good evidence | Weak evidence |
|---|---|---|---|
| Prompt version | Can the agent instructions be repeated? | Prompt, persona, and rule versions are logged. | Instructions change without labels. |
| Rule fit | Did the agent follow the setup and skip rules? | Entries and skips both cite the written rule. | The agent explains decisions after seeing the result. |
| Risk control | Did simulated size, drawdown, and exposure stay inside limits? | Risk behavior is checked before return. | The agent celebrates paper PnL from rule breaks. |
| Journal output | Can another reviewer audit the decision? | Thesis, invalidation, context, result, and next action are visible. | Notes are vague, emotional, or missing skips. |
| Benchmark fit | Was the sample compared against the same review standard? | The benchmark window is stable and comparable. | Every result is judged with a different rubric. |
Example paper-agent evaluation
Prompt: The agent is asked to review BTC and ETH only, use a trend-continuation setup, cap simulated risk, and write a decision note with thesis, invalidation, risk, and skip reason.
Paper sample: Over three weeks, it records 19 simulated entries and 11 skips. The paper result is mildly positive, but the more useful finding is that six skips prevented late entries after the setup had already moved.
Issue: Two entries used unclear invalidation and one entry exceeded the paper risk cap. The agent did not need a new strategy. It needed a stricter output field that rejects any entry without invalidation.
Decision: Keep the agent in paper mode, update one prompt field, and run the same benchmark again. The review uses the benchmark review worksheet instead of treating the first positive sample as proof.
Review questions
- Did the agent know when not to trade? Skips matter as much as entries.
- Did the agent respect the paper risk cap? Rule breaks should not be hidden by positive paper PnL.
- Did the prompt make the output auditable? Missing fields create weak evidence.
- Did the review change only one thing? Small changes make the next paper sample comparable.
How to keep the agent useful without live execution
Keep the agent tied to a written paper-trading system. The paper trading hub should own the overall process. The agent use-case page should define the product frame. The prompt template should turn the rule into instructions. The risk-control page should define stop conditions. The journal should preserve every entry, skip, and review note.
When the agent fails, do not immediately rewrite everything. If the issue is unclear invalidation, fix invalidation. If the issue is late entry, add a timeout or confirmation rule. If the issue is risk behavior, tighten simulated size or pause logic. The point of paper mode is to make those decisions cheaper and more reviewable before any outside live-capital question appears.
AI trading agent without live execution FAQ
How do you evaluate an AI trading agent without live execution?
Evaluate it by versioning prompts and rules, keeping decisions simulated, recording every paper entry and skip, checking risk behavior, and reviewing results before changing the agent.
Can an AI trading agent place live trades in this workflow?
No. This Trading Boy workflow is for paper trading, simulated review, and journaling. It does not execute live trades or hold funds.
What should be versioned when testing an AI trading agent?
Version the prompt, persona, market universe, setup rules, risk limits, model notes, and review date so paper results can be compared fairly.