PABLO-ve: A Component Vocabulary for Interpreting the Behavior of LLM Negotiation Agents

An LLM negotiation agent leaves behind a rich behavioral trace (what it read, what it offered, what it conceded, what it said and why). Saying what the agent did, precisely enough to compare two agents or to locate where a behavior comes from is problematic because there is no agreed upon vocabulary for decomposing the agent and different studies use different evaluation criteria further complicating comparison. In this paper, we extend the BOA (Bidding, Opponent Modeling, Acceptance) agent decomposition with Perception and Language, plus two optional components, Validation and Ending (PABLO-ve). Several published LLM negotiation methods, and all 24 independently authored agents of the ANAC 2026 human-agent league, decompose into these seven components. This decomposition has two main benefits: it allows us to compare methods in a common vocabulary and it allows us to create new methods by recombining components from existing ones. We also define 15 behavioral criteria distilled from the negotiation-evaluation literature (concession dynamics, exploitability, information leakage, strategic coupling and instruction-following alongside deal rate), and run 14 methods, re-expressed as PABLO-ve configurations, over ten scenarios and two backend models (3640 negotiations). The decomposition turns out to be the level at which behavior separates: which component owns the choice of offer splits the fourteen methods into non-overlapping bands on advantage, win rate and Pareto optimality, on both models. An attribution stated in this vocabulary can then be tested by removing the component it names: a powered ablation that disables the Validation component's admissibility check multiplies the rate at which an agent violates its own stated commitments by 4.81, allowing us to localize and interpret the violation.