RSS feedarxiv.org
Symbolic Guidance Improves LLM Agent Coordination in Multiagent SystemsLarge language models are increasingly used as autonomous agents in multiagent systems, but their reliability in distributed coordination remains unclear. Using the AgentsNet coordination benchmark, the study examines whether symbolic guidance derived from established algorithms can regulate agent autonomy and improve performance. It introduces the Symbolic Guidance Taxonomy, which spans open-ended natural-language reasoning, partial pseudocode guidance, and fully prescribed algorithmic execution. The results show that intermediate autonomy levels consistently outperform both unguided agents and fully prescriptive specifications. The authors identify autonomy regulation as a key design principle for LLM-based distributed coordination.