The image shows the process flow for a multi-agent fraud detection system.

• Trigger Event: The flow begins when either the system detects suspicious activity (such as an anomalous transaction) or an analyst issues a query. This event triggers the MCP orchestrator agent.
• MCP orchestrator server: The MCP orchestrator server dispatches sub-tasks to agents, then collects and consolidates results. It acts as a context bridge, translating results from one agent into the formats and requests required by the next. Agents access backend services through the MCP Server’s Tool Box, which can:
  • Query the Oracle Autonomous Database for raw data.
  • Leverage OCI AI Services/LLM for advanced analytics or generative narratives.
• Specialized Agents
  • Data retrieval agent: Upon receiving a request from the orchestrator, this agent queries enterprise sources such as the Oracle Autonomous Database to fetch relevant data such as recent transactions or account histories.
  • Fraud analyzer agent: This agent processes the data (provided by the data retrieval agent) to detect and explain anomalous activity. In its initial phase, it may use rule-based detection or an anomaly detection model. In advanced phases, it integrates OCI Generative AI and LLMs (Large Language Models) by using OCI services.
• Reporting: Once the analysis is completed, results and narratives are routed:
  • To investigators (including comprehensive insights and supporting evidence)
  • To alert and report systems such as mail servers for further notification or escalation