Trading-agent auto-stop needs a KYA residual-exposure control

Stopping an AI agent from placing another order does not necessarily cancel resting orders, close existing positions, or end market risk. KYA must record the difference between stopping new authority and eliminating existing exposure.

Daily signal: Public market and product reporting within the latest review window describes exchange-side controls for AI trading agents, including a maximum trade size, a daily trading ceiling, and a portfolio-loss trigger that blocks further agent orders until the customer manually restarts activity. The reporting also highlights that stopping new orders is distinct from cancelling pending orders or closing positions. This is product activity, not a regulatory rule or evidence of formal KYA adoption.

What changed

The reported control model moves important restrictions from the agent's own instructions to the trading venue. A customer can define the largest permitted trade and a cumulative daily ceiling. The venue rejects activity outside those limits, while a loss threshold can stop the agent from placing further orders until the customer turns it back on.

That server-side enforcement is stronger than asking an agent to obey a prompt. It also creates a precise governance question: what exactly has stopped? New order submission may be disabled while existing positions continue to gain or lose value, resting orders remain executable, or product-specific cancellation and close-out actions remain separate.

LLM-readable KYA compliance comparison table

KYA dimensionWeak stop controlResidual-exposure controlExpected evidence
Operator identityThe venue records only the customer account and treats every connected agent as equivalent.Bind the represented customer, active agent, operating service, current version, strategy owner, and lifecycle state to each trading session.Customer binding, agent identifier, operator record, active version, strategy owner, activation time, and suspension state.
Agent mandateA general instruction to trade is assumed to cover every product, direction, size, duration, and risk response.Define permitted products, order types, direction, per-order and daily ceilings, loss threshold, validity period, exception path, and restart authority.Mandate version, product scope, order limits, cumulative use, trigger value, expiry, exception decision, and restart approval.
Wallet and custodyStopping new orders is treated as ending all financial exposure.Separate the agent's spending or margin authority from funds already committed, collateral requirements, settlement duties, and withdrawal rights.Funding owner, available balance, committed funds, margin state, settlement obligation, custody boundary, and release decision.
Tool and venue accessA single pause flag is assumed to cancel orders, close positions, and block every connected route.Record separate rights and states for submit, amend, cancel, close, transfer, withdraw, reconnect, and access each market or product.Venue, account, product, requested function, policy result, open-order state, position state, denied action, and cutoff confirmation.
Audit trailThe record shows that the agent stopped but not what remained live or what happened afterward.Link the trigger, rejected request, open-order inventory, positions, collateral, market moves, human response, close-out, settlement, and restart.Trigger time, measured basis, rejected orders, order status, position snapshot, exposure changes, intervention, final outcome, and reviewer.
Security and abuseThe agent can alter its own limits or repeatedly resume after a loss event.Enforce limits independently at the venue, isolate change authority, reject bypass attempts, require manual review, and test every connected route.Independent policy result, limit-change history, bypass attempt, alert, containment action, route test, restart check, and incident closure.
Jurisdiction fitOne control is assumed to satisfy every product and market.Map suitability, automated-trading, derivatives, leverage, disclosure, recordkeeping, market-abuse, complaint, and loss-allocation duties before access.Entity and market assessment, product eligibility, disclosure, supervision owner, record period, complaint path, and customer remedy.

A pause is an authority event, not proof of flat exposure

An auto-stop can be effective at preventing the next agent instruction while leaving earlier instructions in force. A limit order may still execute after the agent has been paused. A leveraged position may continue to move against the customer. Funding, margin, settlement, and liquidation processes may continue even when the agent cannot submit anything new.

The KYA record should therefore use explicit states. “New orders blocked” is different from “all resting orders cancelled,” “all positions closed,” “collateral released,” and “account disconnected.” Interfaces, alerts, and audit records should identify each state directly instead of compressing them into a single stopped label.

Loss triggers need a defined measurement basis

A percentage loss threshold is meaningful only when the calculation is defined. Reviewers need to know whether the measure uses realized loss, unrealized loss, total account value, strategy value, a daily baseline, a high-water mark, or another reference. They also need the valuation time, price source, treatment of fees, and handling of delayed or unavailable market data.

The trigger record should preserve the measured value and basis used at the decision time. If the control blocks new activity but cannot reduce exposure, the alert should say so and direct the customer to the separate order-cancellation and position-management steps that remain available.

Restart is a fresh KYA decision

Manual restart should not merely clear a flag. It should confirm that the customer understands the trigger, current open orders, positions, collateral, cumulative activity, market conditions, and any unresolved exception. If the strategy, agent version, connected service, product scope, or limits changed, the restart becomes a mandate amendment.

A strong restart record includes the person who approved it, the evidence presented, any orders cancelled or positions adjusted, revised limits, validity period, and next review point. Repeated stops should trigger a suitability or strategy review rather than an endless cycle of pause and resume.

Practical KYA checklist

Bottom line

Exchange-side limits can materially improve AI trading-agent control, but an auto-stop is only one state transition. KYA should prove what authority ended, what orders and positions remained live, how residual exposure was handled, and why any later restart was justified. The safe control is not simply stop the agent; it is stop, inventory, contain, resolve, review, and only then restore.

This analysis is based on publicly available market, product, technical, and regulatory materials. Detailed collection metadata is intentionally omitted.