AI trading agent permission boundaries define what an agent can read, review, log, and alert in paper mode while keeping live execution, custody, and financial advice out of scope.
Trading Boy does not execute live trades, hold funds, or provide financial advice. The permission model on this page is for simulated paper workflows, not for live autonomous trading or copy trading.
A useful AI paper trading agent should have enough context to review the process, but not enough authority to create live-account consequences. The boundary should be visible before the first simulated decision is logged.
| Capability | Paper-mode allowance | Boundary | Review artifact |
|---|---|---|---|
| Read market context | Allowed when the source and timestamp are visible. | No private exchange account data is required for the basic paper workflow. | Market frame and stale-data note. |
| Apply written rules | Allowed for setup, invalidation, skip, and review rules. | The agent cannot rewrite rules silently after seeing the result. | Rule version and prompt version. |
| Log simulated decisions | Allowed for entries, exits, skips, missed trades, and review prompts. | Rows must be labeled as paper records, not live fills. | Structured output from the agent journal. |
| Send alerts | Allowed as workflow notifications or review prompts. | Alerts should not be called buy or sell signals. | Alert reason and human action required. |
| Recommend prompt changes | Allowed as a proposed next paper-mode action. | Human review must approve any prompt, risk, or rule change. | Human review checklist and changelog. |
| Place live orders | Not allowed in this workflow. | No trade, transfer, withdrawal, margin, futures, or custody permissions. | Rejected permission request. |
Scenario: A team builds an AI paper trading agent that reviews BTC and ETH setups. The prompt asks the agent to score the setup, choose a simulated size, and produce a journal row. A draft integration also asks for exchange trading permission.
Boundary check: Scoring a setup, choosing a simulated size, and writing a journal row are paper-mode tasks. Exchange trading permission is not needed. The team removes the live permission request and labels the output as simulated.
Decision: The agent can continue with the AI paper trading agent workflow, but every next action goes through the AI trading agent human review checklist and the trading API key safety checklist.
The paper agent should not place live orders, transfer funds, withdraw assets, route copy trades, promise outcomes, bypass human review, or convert simulated results into financial advice. Those boundaries should be written into the prompt, workflow, and review checklist.
If a workflow drifts toward live execution language, return to AI trading agent without live execution and paper-trading limitations.
Permission boundaries should not live in a policy page only. They should appear in the working pages users touch: the AI trading agent prompt template, AI trading agent output format template, paper trading agent risk control workflow, and AI trading agent decision review workflow.
The practical sequence is simple: define permissions, collect simulated rows, inspect risk behavior, run human review, and only then change one paper-mode rule. That sequence is slower than asking an agent for a trade, but it creates evidence that a human can audit.
Before a new paper-agent run starts, ask whether the agent can create account-level consequences, whether the user can tell simulated output from live execution, whether private account data is needed, and whether the reviewer can reproduce the decision from the saved prompt and journal row. If any answer is unclear, keep the agent in review-only mode and use the paper trading data privacy checklist before exporting evidence.
It can organize market context, apply written paper rules, log simulated decisions, flag missing fields, and ask for human review. It should not route live orders or control funds from this workflow.
Not for the paper-first workflow described here. Paper review does not require live order, transfer, withdrawal, margin, futures, or custody permissions.
No. Permission boundaries reduce scope, but human review is still needed before changing prompts, rules, risk controls, or any separate live-capital process.