Agent Overview
Korastratum ships 22 autonomous agents organized into two intelligence tiers across 7 product areas. Each agent follows the same Observe-Decide-Act pattern and supports shadow mode for safe evaluation.
Tier 1: Rule-based (Core)
These agents use deterministic rules, threshold logic, and pattern matching. They are fast, predictable, and fully auditable.
| Agent | Product | What it does |
|---|---|---|
| Reconciliation Agent | GL | Analyzes reconciliation breaks and auto-resolves simple ones (timing differences, rounding, duplicates) |
| Settlement Agent | Settlements | Triggers EOD settlement batch generation and monitors the processing lifecycle |
| Elder Detection Agent | Compliance | Monitors elderly customer accounts for exploitation signals — velocity spikes, new beneficiaries, large withdrawals, unusual hours |
| Compliance Monitor Agent | Compliance | Re-screens high-value transactions against updated sanctions/PEP lists via continuous monitoring |
| Branch Ops Agent | Inventory | Monitors inventory levels and auto-generates purchase orders at reorder point |
| Escalation Routing Agent | Compliance | Analyzes compliance alerts, predicts SLA breaches, and routes to the appropriate specialist |
| Treasury Liquidity Agent | Treasury | Monitors NIP Clearing account balance and predicts settlement funding shortfalls |
| Rewards Optimization Agent | Rewards | Detects reward fraud (referral rings, point abuse) and disables rewards for fraudulent accounts |
| Corridor Performance Agent | Corridors | Analyzes remittance corridor performance and recommends fee/capacity adjustments |
| Adaptive Liveness Agent | IDV | Monitors IDV liveness challenge success rates and recommends parameter adjustments (advisory only) |
| Caregiver Monitor Agent | Compliance | Tracks POA/caregiver activity on elderly accounts and alerts family contacts for suspicious patterns |
| Bill Pay Recovery Agent | Payments | Monitors failed bill payments and auto-retries transient failures with exponential backoff |
Tier 2: ML-powered
These agents use Bayesian scoring, anomaly detection, and machine learning models. They require shadow mode validation before activation.
| Agent | Product | What it does |
|---|---|---|
| IDV Decision Agent | IDV | Reviews KoraIDV manual review queue with confidence-weighted scoring to auto-resolve clear approve/reject cases |
| Compliance Disambiguation Agent | Compliance | Cross-references PEP/sanctions matches against KoraIDV document data to reduce false positives (40-50% reduction target) |
| Lending Decision Agent | Lending | Pre-screens loan applications using credit scoring, transaction analysis, and DTI calculations |
| Quote Optimization Agent | Corridors | Analyzes FX rate patterns to suggest optimal transfer timing and auto-refresh expiring quotes |
| Name Matching Agent | IDV | Intelligent name comparison with OCR error correction and cultural naming convention handling (15-20% fewer false rejections) |
| Case Investigation Agent | Compliance | Auto-investigates CEngine compliance cases in ReviewRequired status with Bayesian confidence scoring (50-70% auto-resolution) |
| Match Verification Agent | Compliance | Verifies screening matches using field-level Bayesian scoring with cultural name awareness (40-50% false positive reduction) |
| Predictive Risk Agent | Risk | Builds behavioral baselines and detects anomalies using standard deviation analysis before they become violations |
| Fraud Prevention Agent | Fraud | Real-time fraud pattern detection (ATO, structuring, velocity breach) with autonomous freeze/OTP/flag response |
| Fraud Ring Detection Agent | Fraud | Clusters face embedding duplicates with temporal/geographic context to detect fraud rings vs legitimate retries |
Agent lifecycle
Every agent goes through the same lifecycle:
- Registered — The agent type exists in the platform registry
- Enabled (Shadow) — The agent is enabled for a tenant in shadow mode; it observes and decides but does not act
- Observation period — Minimum 2 weeks of shadow decisions are compared against human outcomes
- Activation review — If the agreement rate meets threshold, the agent can be promoted
- Enabled (Active) — The agent acts autonomously within its risk threshold; decisions above threshold are escalated to humans
See Shadow Mode for details on the observation period and activation criteria.