Amazon Science showed agentic AI cut root‑cause analysis to minutes
The system identified network faults in minutes during a trial with NTT DOCOMO at MWC.
Amazon Science released a blog post on Oct 1 2026 describing how an agentic AI system analyzed a telecom network in minutes.
The system builds a continuously updated “digital twin” – a graph where every device and link is a node or edge, enriched with alarms and performance metrics.
The agentic layer selects the appropriate graph algorithm for each incident.
When a failure occurs, the AI runs a three‑stage cascade. First it isolates the most‑connected sub‑graph, then groups nodes with community‑detection, and finally ranks candidates with customized centrality measures that focus only on alarming nodes.
Personalized PageRank starts random walks from the devices that raised alerts, degree centrality counts only connections to those alerts, and closeness measures distance to them. "The agentic layer selects the appropriate graph algorithm for each incident." The agent decides which mix fits the observed topology.
In a trial with NTT DOCOMO at the Mobile World Conference, the pipeline identified the root cause in a matter of minutes instead of hours.
After ranking, the agent expands a small failure sub‑graph, checks historical incidents and runbooks, and assigns a confidence score before opening or updating a trouble ticket.
An on‑demand AI assistant lets engineers query the twin, test alternative hypotheses, or request deeper analysis at any stage. Think of it as a network‑savvy detective who never sleeps.
The approach can be extended with graph neural networks or graduated autonomy, but the current version does not yet perform fully automated remediation or learn new skills without human approval.
Source: Amazon Science