Confucius: Intent-Driven Network Management with Multi-Agent LLMs#

Zhaodong Wang et al. (Meta, JHU, Harvard, Stony Brook)

Why: First credible hyperscale report on multi-agent LLMs for network ops — useful signal on agentic operations adjacent to the PM's agentic-identity interest.

What: A production multi-agent framework, operational two years with 60+ apps, modeling management workflows as DAGs.

How: LLMs integrated with existing tools, RAG for long-term memory, human/model interaction primitives, and tight coupling to existing validation to prevent regressions.

Where applicable: Meta-scale ops; the architecture is a reference point for agentic network management generally.