Greenlight Financial says a multi-agent system now resolves more than 90% of its monthly security and risk cases without human intervention. The company receives roughly 3,000 to 4,000 alerts a month, covering compromised cards, access anomalies, infrastructure warnings and suspicious financial activity.
Its Luna system runs on Amazon Bedrock AgentCore and uses more than 30 specialized agents. A first-tier supervisor gathers evidence and handles routine cases in seconds. Cases below its confidence threshold move to a second tier that correlates three to five data sources and prepares a detailed report for human review.
A 30-day episodic memory helps identify recurring patterns and deduplicate alerts. Longer-lived rules and resolution templates preserve institutional knowledge. Greenlight says it evaluates the system against labeled cases and routes unfamiliar or high-impact decisions to analysts rather than allowing the agents to act without limits.
The performance figures come from Greenlight and AWS, not an independent audit. Even so, the architecture illustrates a practical boundary for autonomous security work: automate repetitive investigation, preserve an evidence trail, and keep people responsible for uncertain or consequential cases.