From Moses to Multi-Agent Systems: A Paradigm Shift in Machine Translation

Machine translation has been going through several significant phases, ranging from rule-based to statistical to neural networks and Transformers, large language models (LLMs), to agentic and multi-agent systems. New capabilities have been introduced with each stage, including improved representation, handling of wider contexts, following instructions, access to external knowledge, and iterative verification. In existing literature, these developments are investigated individually, though, one may have a better idea of the capacities of translation systems by analyzing these transitions. This paper introduces a capability-based framework (the Translation Intelligence Continuum) to analyze the evolution of machine translation that goes beyond the focus on the machine translation architecture itself. It covers prediction and representation, context, generalisation, following instructions, interaction, contextual inference, using tools, verification, autonomy and collaboration. Within this context we discuss each major transition: what were the shortfalls of the previous approaches, what were the innovations of the new approaches, what new capabilities were created, and what challenges were left. We also propose a conceptual architecture for agentic and multi-agent translation that integrates planning, specialised agents, retrieval, criticism, and verification. This work provides a unified perspective on the evolution of machine translation and highlights the trade-offs between increasing system capabilities, translation quality, computational cost, and system complexity.