Contents

06Architecture

Agent Intelligence and Strategy

The intelligence layer proposes. The policy layer decides. Nothing executes because a model asked.

The agent system separates interpretation from settlement. The intelligence layer analyses data and generates plans. The account cannot execute a transaction merely because a language model requested one.

Every strategy passes deterministic policy validation first.

From instruction to executable policy

The instruction “trade my $10 toward $100, but stop at $5” is not a complete transaction policy. The application asks for, or applies disclosed defaults for, the maximum allocation per trade, eligible assets, permitted venues, strategy term, borrow permissions, pricing method and time-out rules.

Policy elementIllustrative settingPurpose
Initial capital10 USDCStarting account funding
Goal threshold100 USDCReference point for the target state
Risk threshold5 USDCTrigger for a new-risk halt
Per-order exposure cap20% of eligible equityLimit individual actions
BorrowingDisabledAvoid unintended leverage
Authorised assetsOwner-selected allowlistRestrict asset exposure
Session duration24 hoursBound credential lifetime

Strategy states

StateMeaningPermitted behaviour
DraftUnconfirmed strategyNo execution
AuthorisedOwner-approved policyMay begin validated execution
ActivePermission valid and agent runningApproved actions
PausedNew execution suspendedOwner controls and safe unwinds
RestrictedGuardrail violation or external problemNo new risky actions
ExpiredSession or policy window endedRequires reauthorisation
SettlingClosing open positions and obligationsNarrow settlement actions only

Model separation

Data collection, research, plan construction, risk validation, signature generation and submission are independently observable subsystems. The model's narrative explanation is useful for transparency. It is not a substitute for deterministic transaction validation.