Graph-centric agentic intelligence

Amazon Science published a post describing a graph-centric agentic AI approach for network root cause analysis, combining cascaded graph algorithms with agentic-AI execution. The work was demonstrated with NTT DOCOMO at the Mobile World Conference (MWC) earlier this year, achieving root cause analysis in minutes on commercial networks, according to the post.

The approach rests on three pillars. First, graph modeling: the network is represented as a continuously synchronized graph whose vertices are devices with attributes and whose edges represent connections. This "digital twin" ingests network dependencies, live alarms, and key performance indicators (KPIs) from multiple data sources across network segments and layers.

Second, graph-centric analytics: a three-stage cascade narrows the search space. Stage 1, decomposition, identifies the most-connected parts of the topology and localizes analysis to boundary nodes, narrowing candidates from thousands of nodes to hundreds. Stage 2, clustering, applies community detection algorithms (Louvain or label propagation) to narrow candidates from hundreds to tens. Stage 3, centrality ranking, applies personalized PageRank, alarm-relative degree centrality, and alarm-relative closeness, each conditioned on the set of alarming nodes.

Third, graph-based agentic orchestration: AI agents select and compose graph algorithms adaptively. Agents first query the incident knowledge base and apply a matching prescribed remedy directly if one exists at high confidence; otherwise they invoke the full cascade after a complexity triage. Results feed a failure subgraph, a confidence-scored root cause determination, and either a new or existing trouble ticket, with NOC feedback for continuous learning. An on-demand AI assistant lets engineers query the digital twin interactively.

The post traces prior graph research in networks: topology graphs, knowledge graphs and ontologies adding semantics, alarm correlation graphs requiring predefined hierarchies, dependency graphs autogenerated in real time from SDN and NFV controllers enabling Bayesian fault localization at 95% accuracy in under 30 seconds with no manually authored rules, causal subgraphs from live alarms converted into directed acyclic graphs, and graph neural networks. It also notes that remediation in traditional network operations centers for complex multilayer failures averages four to five hours and can extend to days.

Stated future directions include graduated autonomy, where autonomy is earned rather than granted by default, and self-learning agents that propose repeatable resolution procedures as candidate skills requiring explicit user approval, with the user retaining full governance.