Discovering Agent Strategies with ATLAS

Summary & Methodology Analysis

Building production systems with LLM-based agents often hits a wall when developers need to understand, monitor, and analyze complex decision-making strategies. To solve this, the paper presents ATLAS, a framework that extracts explicit and human-interpretable behavioral models from agent runs. The process begins with trace collection, which records raw agent-environment interaction traces. Next, trace abstraction uses an LLM to categorize raw actions and observations into semantic high-level symbols through categorization, category normalization, and per-trace assignment. Once abstracted, automata learning applies the Alergia state-merging algorithm to infer probabilistic finite-state Markov chains from these traces. These learned models are then leveraged via model-driven engineering for downstream tasks like explanation, symbolic knowledge transfer, and formal analysis.

In practice, these learned Markov chains unlock powerful capabilities. For instance, the dynamic knowledge transfer mode enabled a 14B parameter model, ministral-14b, to successfully solve 38.3% of penetration-testing benchmark tasks. Additionally, model engineering can significantly simplify the resulting graphs. In one case study, model slicing and pruning successfully reduced a Markov chain from 41 states to 7 states, making the underlying strategy much easier to inspect.

However, the approach introduces specific constraints that engineers must consider. The abstraction process for large or diverse datasets may encounter scalability issues. Furthermore, the pruning method used for explaining agent behavior can inadvertently collapse important decision-making steps, meaning automated simplification carries a risk of losing critical execution details. The paper does not specify hardware requirements, runtime latencies, or token costs.