Strategy guide
Audit what your AI trading agent actually did
An AI trading audit starts with account records, not the assistant’s summary. A prompt is an instruction, a tool response may show only a submitted action, an order can remain open or fill partly, and a position can change later through another order, funding payment or liquidation. The account’s orders, fills, positions and ledger are the evidence.
Fensory exposes read tools for account state, spot state, open orders, order status, fills, funding payments and portfolio history. An assistant can use them to reconstruct what was requested, what was submitted, what actually filled and what remains exposed. This guide stays read-only and produces an evidence table rather than a performance claim.
Copy this prompt
Use only read tools to audit my Fensory trading activity from [START TIME AND TIMEZONE] to [END TIME AND TIMEZONE]. Call get_account_state, get_spot_account_state, get_open_orders, get_fills, get_funding_payments and get_portfolio_history. Use get_order_status for every order ID I provide or every relevant order ID returned by the records.
Build a chronological table with timestamp, instrument, spot or perpetual-futures surface, requested action if supplied, submitted order, status, filled size, average fill price, fee, funding payment, resulting position and evidence source. Keep “prepared”, “submitted”, “partly filled”, “filled”, “cancelled”, “rejected” and “still open” as different states.
Reconcile opening and closing account state. Flag unmatched fills, duplicate orders, open remainders, changes in leverage or margin mode, and any market outside the intended allow-list. Calculate realized trading profit or loss only when the returned records support it. Report unrealized profit or loss separately. Do not infer that the agent caused a manual trade, deposit, withdrawal or account change unless the evidence identifies that source.
Do not place, cancel, replace or close any order. Do not change leverage, limits or account settings. If the returned history is incomplete, say which record is missing and stop short of a performance conclusion.Who this is for
This guide is for a user, operator or reviewer who needs to verify an agent trading session, investigate an unexpected position or build a dated record before evaluating a strategy.
What it is not
It is not an audited financial statement, tax record, proof that a prompt caused every account event or a public performance leaderboard. It does not infer intent from a fill and does not label an unrealized gain as realized profit.
What you need
A defined time window
Record the start and end time with a timezone. Without a fixed window, earlier positions and later manual actions can be mixed into the agent session.
The account and market scope
Name the signed-in account, spot or perpetual-futures surface and every market the agent was allowed to use. A conversation alone is not proof of which account a tool read.
The original instructions
Keep the prompt or task record as context, then test it against account evidence. The audit should be able to say that an instruction was not executed or that an execution had no matching instruction in the supplied record.
What the agent does
01
Freeze the scope
The audit defines account, markets, time window and supplied instruction record before reading results. That prevents the reviewer from selecting only favorable fills or attributing unrelated activity to the agent.
02
Read opening and current state
Positions and balances provide the boundary conditions. A current snapshot cannot reconstruct every past state, so the output labels what is observed now and what comes from historical records.
Tools: get_account_state · get_spot_account_state
03
Collect orders and fills
The assistant separates order submission from execution and records partial fills and resting remainders. An order ID links the status response to the fill evidence without relying on conversational memory.
Tools: get_open_orders · get_order_status · get_fills
04
Add funding, fees and portfolio history
Funding and fees can change the result after entry, while portfolio history shows account-level movement. The audit reports these components separately before calculating any supported total.
Tools: get_funding_payments · get_portfolio_history
05
Reconcile and flag gaps
The final table explains every supported state change and lists records that do not reconcile. Missing history remains missing rather than being filled with an assumption about what the agent intended.
Failure modes and limits
A tool call is mistaken for a fill
A successful submission can leave an order open, partly filled or rejected later. Order status and fills decide what executed.
The conversation is treated as the ledger
An assistant can summarize incorrectly or lose context. Use the prompt as intent evidence and the account records as execution evidence.
Manual and agent actions are mixed
The same account can contain activity from another client or a person. When source attribution is unavailable, the audit says the cause is unknown rather than assigning it to the agent.
Open risk is omitted
A positive realized fill result can sit beside a losing open position or resting order. Current positions and open orders remain in the report after the closed-trade calculation.
Funding and fees disappear
Gross price change is not the account result. The audit includes trading fees and funding payments and says when a record is unavailable.
Read the result in the terminal
- Orders
- Use order IDs, status, original size, remaining size, order type and timestamps to identify what was submitted and what remained active.
- Fills
- Use filled size, price, side, fee and timestamp as execution evidence. Several fills can belong to one order, and one partial fill does not close the order.
- Positions and ledger
- Reconcile current size, entry, mark, realized result, unrealized result and funding. Keep deposits, withdrawals and transfers outside the agent-performance calculation unless the task specifically includes them.
Runs on
The client list comes from the same release record as the MCP page. Every verified client has its own setup guide.
Questions people ask
- What proves that an AI agent placed a trade?
- The strongest evidence is an order record linked to fills and the resulting account state. A prompt or assistant message alone proves only what was requested or reported.
- Is a submitted order a completed trade?
- No. It may remain open, fill partly, be cancelled or be rejected. Check status and fills before calling it executed.
- Can the audit calculate profit?
- Only when the returned records support entry, exit, fees and funding for the defined window. Unrealized results and unmatched positions remain separate.
- Can the audit tell whether a human or agent caused every action?
- Not always. If the source is not present in the records, the audit should label it unknown rather than infer attribution from timing.
- Does this audit change the account?
- No. The workflow uses read tools only and explicitly forbids order, leverage and account-setting changes.