Teaching AI to Fight Fires: Autonomous Reliability Agents at Meta Scale#

Gaurav Mitra (Meta)

• Why: The most quantitatively substantiated AIOps deployment in the program: an investigation agent that has handled 1,000+ real incidents, cut detection-to-mitigation time by 60%, and matches senior on-call engineers ~80% of the time. That is deployment evidence, not a demo — and it is squarely in the "ML for infrastructure operations" interest area.
• What: The "reliability flywheel": an always-on investigation copilot that encodes senior on-call engineers' investigation methods and improves with every incident, deployed across Meta's recommendation systems. The talk also lays out a staged autonomy progression: pattern identification, autonomous investigation, supervised mitigation of low-risk reversible actions, and self-healing as the end state — plus an explicit treatment of the trust boundary between agents and humans. (from abstract only; the event page's article teaser adds that the originating December 2024 config change caused a widespread routing error across every region, taking 50+ engineers and 28 hours to recover)
• How: Context-engineered architecture correlating time-series metrics, deployment logs, configuration changes, and infrastructure health, built on structured tool servers and reusable workflows rather than free-form prompting.
• Where applicable: Validated on Meta's recommendation infra (org-specific numbers), but the architecture — tool servers + reusable investigation workflows + human-in-the-loop trust ladder — is a transferable pattern for incident automation in any environment.