Root‑cause analysis (RCA) in large‑scale distributed systems is difficult because a single symptom can span services, data, configuration, compute and organizational boundaries. Responders must piece together incomplete evidence and coordinate experts who own different parts of the causal chain. Generic language‑model agents that rely only on tool access fail to address these issues: they may use stale or mistimed context, treat failed queries as missing symptoms, report unsupported system states, and overload observability services built for human‑paced usage.
Diviner (Domain‑Informed Verifiable Incident Navigation through Evidence and Review) is an agentic RCA system deployed at Meta, organized around a four‑stage reliability lifecycle.
In the Encode stage, a large corpus of reviewed incidents is combined with service‑owner expertise to produce scoped, versioned skills, workflows, typed tools and an incident knowledge graph.
During Execute, an adaptive Lead and context‑isolated Workers apply the relevant knowledge to point‑in‑time evidence while maintaining evidence‑linked hypotheses.
Evaluate compares saved investigation trajectories with responder‑reviewed outcomes, measuring exact cause, useful direction and Groundedness.
Adapt uses recurring knowledge gaps and on‑call feedback to derive focused operational‑knowledge updates, admitting them only after replay, quality checks and owner approval.
By tightly coupling human governance with autonomous agents, Diviner enables fast fault localization, verifiable reasoning and continuous knowledge evolution for large‑scale infrastructure, markedly improving RCA efficiency and reliability.
Review