ATLAS reverse-engineers AI agent strategies into interpretable models

As LLM-based agents take on complex tasks like software testing and cybersecurity assessment, their behavior remains frustratingly opaque. Existing evaluations focus on whether tasks succeed or on raw execution traces, offering little insight into the strategies actually employed. ATLAS (Automata Learning for Agent Trajectory Analysis and Strategy Discovery) addresses this by combining trace abstraction with automata learning to infer finite-state models that capture agent-environment interaction strategies. These models are human-interpretable and allow automated analysis of recurring behaviors, decision points, successful completion paths, and failure loops.

In a proof of concept, the researchers applied ATLAS to trajectories from an LLM-based penetration-testing agent targeting 12 vulnerable machines. The learned models exposed high-level behavioral strategies for exploitation that are difficult to spot in raw traces. The team also demonstrated symbolic model-based knowledge transfer, using the discovered behavioral models to distill strategies from powerful frontier models into compact language models. Accepted at ACM/IEEE MODELS 2026, ATLAS opens a new path for model-driven engineering: transforming opaque agent trajectories into explicit behavioral models that support explainability, auditing, and systematic analysis of agentic systems.