Agentic MT frameworks often adopt human-inspired roles, layered post-editing, and complex workflows. To stress-test these designs, we examine cost–performance trade-offs across five multi-agent MT workflows on specialized long documents. Using six language pairs from two financial test sets and LLMs spanning lower to higher capacity and cost tiers, we evaluate translation quality and terminology accuracy relative to token-level cost. We find: (i) lean pipelines match the quality of more complex systems at lower cost; (ii) terminology accuracy depends more on model capacity than on workflow complexity; (iii) document-length resilience is not universal, as some simple and some complex workflows maintain quality on longer documents; and (iv) post-editing agents offer a better return on model capacity than base translation. Edited-span analysis further suggests that many post-edits are stylistic or lexical alternatives rather than consistently preferred improvements.

For language models to generalize correctly to novel expressions, it is critical that they exploit access compositional meanings when this is justified. Even if we don’t know what a “pelp” is, we can use our knowledge of numbers to understand that “ten pelps” makes more pelps than “two pelps”. Static word embeddings such as Word2vec made strong, indeed excessive, claims about compositionality. The SOTA generative, transformer models and graph models, however, go too far in the other direction by providing no real limits on shifts in meaning due to context. To quantify the additive compositionality, we formalize a two-step, generalized evaluation that (i) measures the linearity between known entity attributes and their embeddings via canonical correlation analysis, and (ii) evaluates additive generalization by reconstructing embeddings for unseen attribute combinations and checking reconstruction metrics such as L2 loss, cosine similarity, and retrieval accuracy. These metrics also capture failure cases where linear composition breaks down. Sentences, knowledge graphs, and word embeddings are evaluated and tracked the compositionality across all layers and training stages. Stronger compositional signals are observed in later training stages across data modalities, and in deeper layers of the transformer-based model before a decline at the top layer. Code will be publicly available on GitHub upon acceptance.