Pradipta Ghosh, Gaurav Mitra, Shripad Nadgowda, Akshil Shah, Junyu Liu, Canturk Isci, Syamla Bandla, Xinyao Hu, Onkar Deshpande, Anik Dasgupta
A governed multi-agent RCA system deployed at Meta reduced detection-to-mitigation time by about 52% and operational effort by about 15% on onboarded incident cohorts, while identifying the exact root cause in 45% of scored investigations.
Observational cohort comparisons of Diviner-assisted vs unassisted incidents across Meta's recommendation and ranking infrastructure, plus scored evaluations of saved trajectories against responder-reviewed outcomes, on a corpus of several thousand incidents including several hundred high-severity ones, with several thousand alert-triggered investigations per month.
Approximately 52% lower detection-to-mitigation time and 15% lower pre-mitigation effort in assisted cohorts; exact cause 45%, useful causal direction 60%, mean Groundedness 80%; first useful medium-confidence hypothesis at 7.0 min p50 / 13.7 min p95, final conclusion ~15 min p50 / 24.0 min p95. Effects hold only within Meta's onboarded, governed-knowledge environment; no external system replication is shown.
The headline outcome figures are observational cohort associations, not randomized assignments, so selection bias (which incidents got the agent) is unaddressed in the provided text; results are single-organization and tied to Meta's bespoke tooling and incident corpus; judge-based scoring (DivinerEval) is validated internally; no code or benchmark release evident.
Cohort assignment is likely non-random: simpler or recurring incidents may be the ones the agent handles, inflating the 52%; 'useful direction' is a generously defined success metric; the heavy infrastructure of governed skills and human review is the real driver, not model capability, so a generic-agent baseline comparison (mentioned but details truncated) is the key number and may be modest; industry lab PR incentive applies.
Concrete, if modest: agent systems with heavy governance scaffolding can deliver measurable operational savings in large engineering organizations within a year, supporting the view that AI ROI shows up first in internal infrastructure and SRE functions rather than end-user products; nothing here speaks to external revenue, capex, or labor markets yet.
Labor & Knowledge Work: Suggests SRE/on-call work is being reshaped toward review and governance rather than manual investigation, an early concrete case of AI shifting (not eliminating) expert roles.
Technical Capability: Evidence that agentic reliability in high-stakes domains comes from engineered governance (evidence semantics, human-gated adaptation), not raw model gains.
Market & Valuation: Marginally relevant: shows enterprise AI value accrues as internal productivity, hard to monetize externally; weak signal otherwise.