The field of multi-agent LLM systems is rapidly evolving, making it challenging to keep up with the latest research. The author proposes a framework to understand the field’s development, dividing it into two waves: the first wave (2023) focused on whether multiple LLMs could coordinate at all, while the second wave (2025 and beyond) is concerned with understanding why these systems fail and how to improve them. The emergence of agentic coding systems in 2024, which demonstrated the effectiveness of single agents with well-designed tool interfaces, significantly impacted the field and narrowed the scope of multi-agent systems.

The Wave 2 timeline illustrates a shift towards measuring, taxonomizing, and fault-injecting multi-agent systems (MAS). Papers like MAST, MAS-FIRE, and Silo-Bench highlight system design flaws, the capability paradox, and the communication-reasoning gap, respectively. This wave emphasizes rigorous testing and verification, moving beyond the initial trust in role structure and dialogue.