TRACE: A Multi-Agent System for Natural Language-Driven Social Graph Investigation
Arunachaleshwar Ravichandran (Meta), Nicole Chen (Meta), Ankitesh Gupta (Meta), Antonios Broumas (Meta), Ioannis Konstantakopoulos (Meta), Seyoung Park (Meta)
Architectural Patterns & Composition
Summary
A multi-agent system for social graph forensics that uses natural-language behavior detection and LLM-driven graph exploration, achieving 10x network expansion and 91.9% discovery of unknown suspicious entities.
Description
Trust & Safety teams at social platforms must investigate complex networks of suspicious entities — groups, pages, and profiles — that span multiple hops across the social graph. Traditional approaches rely on hard-coded rules or single-entity classifiers that cannot perform multi-hop reasoning or adapt to novel attack patterns without code changes. We present TRACE, a multi-agent LLM-powered system for semi-automated social graph forensics. TRACE introduces two key innovations: (1) natural language behavior detection, where suspicious patterns are defined in prose rather than code and an LLM agent translates these into executable data queries at runtime; and (2) LLM-driven graph exploration, where the system performs weighted breadth-first traversal of the social graph, with the LLM assigning information gain scores to candidate edges to decide which entity relationships to follow. Multiple specialized agents — each behavior spawning its own dedicated Specialist — collaborate through a deterministic pipeline where LLM calls are confined to agent boundaries and all control flow is ordinary code. In evaluations on two real-world investigations, TRACE achieves a 10× network expansion factor, discovers 91.9% previously unknown suspicious entities, and identifies behavioral signals with 2.2× lift over content-only analysis.
