Agentic Trading
SAE — Survivability-Aware Execution + 5-tier liquidity. How TRUE keeps autonomous logic survivable.
What this is
When agents stop being assistants and start being actors, the rules have to change. Agentic Trading is opt-in and bounded: every autonomous action runs through SAE, lives under user-set caps, and emits a notification with the route plan. This page is how TRUE keeps autonomous logic survivable.
Survivability-Aware Execution (SAE)
SAE is a middleware layer that sits between the strategy engine and the exchange executor. Its job is to treat every upstream intent — even from our own agents — as untrusted by default. Published as an arXiv paper in March 2026, it is now in production on every agentic execution path.
Exposure Budgeting
Hard caps on per-asset, per-strategy, and aggregate exposure. Budgets reset on cooldowns, not on PnL.
Trust-State Conditioning
Agents earn execution latitude through clean track records and lose it on anomalies — instantly, not retroactively.
Cooldowns & Allowlists
New venues, new pairs, and new strategies enter through allowlist gates. Cooldowns prevent reentrant flurries.
Staged Execution
Larger sizes are split into stages, each gated by post-fill checks against expected impact.
The Delegation Gap
SAE formalizes a metric called the Delegation Gap — the distance between what an agent intended and what would actually have happened on the market without protection. Backtests on Binance replay data (BTCUSDT/ETHUSDT perps, Sep–Dec 2025) showed substantial drawdown and tail-loss reduction with SAE active. The metric is computed per-strategy and per-agent and is visible in the audit trail.
Five-tier liquidity architecture
Not every asset deserves the same execution model. TRUE’s liquidity stack is split across five tiers, each with its own A-Book / B-Book balance.
- Opt-in per user, with hard exposure caps you set yourself.
- Every autonomous action emits a notification with the route plan and the trigger reason.
- Pause and revoke are one-click and instant — across the orchestrator and the executor.
- An agent cannot raise its own exposure cap. Cap changes require a fresh user signature.
- Strategies that breach a cap get a 24h cooldown automatically. Repeated breaches escalate to a manual review.
- Anomaly = revocation. An agent that violates trust-state thresholds is paused immediately, not retroactively reviewed.
Strategy SDK
The strategy SDK ships in TypeScript and Python. A strategy is a pure function from observation → intent; the SDK runtime serializes the intent into the SAE wire format and submits it for evaluation. SAE returns a verdict (allow, partial, deny, cooldown) plus a reasoned justification.
import { defineStrategy, type Observation, type Intent } from '@true/strategy-sdk';
export default defineStrategy({
name: 'sol-mean-reversion',
schema: { /* zod schemas for state, params */ },
async tick(obs: Observation, ctx): Promise<Intent | null> {
if (!obs.signal || obs.signal.conviction < 0.7) return null;
return ctx.intent({
side: 'long',
asset: 'SOL',
size_quote_usd: Math.min(ctx.budget.remaining, 250),
stages: 2,
reason: `signal-${obs.signal.id}`,
});
},
});Simulator and paper-trading mode
Every strategy runs against the historical replay simulator before it can graduate to production. The simulator uses real TRUE Quotes snapshots and venue order-book replays, applies SAE verdicts deterministically, and produces a Delegation Gap distribution. Paper-trading mode runs the same code against live data without submitting orders — same telemetry, no exposure.
Audit trail format
Every executed intent is appended to a per-user audit log with: trigger reason, agent identity, SAE verdict, route plan, fill outcome, and post-fill impact vs expected. The log is exportable as CSV or fetched via the audit API. Pull it before you investigate anything.
Offering agentic execution to your users requires:
- Surfacing the not financial advice notice on every agentic surface.
- Geo-blocking restricted jurisdictions before the user can opt in.
- Storing the audit trail for the period required by your jurisdiction.
- Providing a one-click revoke that propagates to TRUE within seconds.
Required disclosures, contractual minimums, and onboarding flow on Partner Integration.
Safety, limits, failure modes
- What signals an agent has gone off-policy: repeated denied intents, a sudden burst of intents on a new venue, intents that don’t reference a recent observation. The audit trail surfaces all three.
- Cooldowns compound. Two breaches in a 24-hour window double the cooldown.
- Partial fills. Staged execution can produce a partial fill if the second stage fails the post-fill impact check; the residual is cancelled, not retried.
- Trust-state recovery. Trust state can be re-earned with a clean run, but the cooldown floor never disappears entirely for a previously-anomalous agent.
See also
- Signals — what triggers most agentic intents.
- Swaps — the execution path SAE wraps.
- Safety Overview — the broader safety model.