🤖 LLM Agents
Agent Architecture
| Agent | LLM | Role | Context |
|---|---|---|---|
| TriageAgent | Gemini 3 | Severity re-classification from tool output | 1M tokens |
| CorrelatorAgent | Claude | Cross-finding deduplication & linking | 200K |
| AttackChainAgent | Gemini 3 | Multi-step exploit path detection via DAG | 1M tokens |
| RemediationAgent | GPT-4 | Fix suggestions with code patches | 128K |
| SummaryAgent | Claude | Executive summary generation | 200K |
| ExploitSkeletonGen | Gemini 3 | Initial PoC skeleton for GAEV | 1M tokens |
Semantic Refinement Pattern
The core principle: "Tools detect, AI understands." Traditional tools run in parallel, their outputs are normalized into a unified schema, then 3 LLM models vote on severity re-classification. Result: -54% false positives, F1 = 0.86 across 53 contracts.