Self-Evolving Agent
Hierarchical multi-agent system that runs the threat-to-patch loop autonomously.
A hierarchical, self-evolving agent system that runs the threat-to-patch loop autonomously across multiple Amazon AI services. Three top-level agents (Triage, Red Team, and Model Patch & Handoff) coordinate specialized subagents:
- Triage detects emerging threats from tickets and logs and routes them to the right defender.
- Red Team discovers and validates new attacks autonomously, building a continually refreshed adversarial dataset.
- Model Patch & Handoff applies continual fine-tuning, runs automated benchmarks, and gates the handoff to deployment.
The interesting part is the design pattern: separating threat discovery from defense generation lets each side improve without destabilizing the other. The system compresses launch cadence by ~75% with limited human-in-the-loop oversight, turning what was a months-long model update cycle into weeks.
The lesson worth porting elsewhere: the bottleneck is rarely model quality, it is the speed at which the threat-fix-ship loop closes. Autonomy across that whole loop, not just one stage, is what compresses time.