Our Journey to Safely Unleash Agents at Meta Scale#
@Scale: Systems & Reliability 2026 digest Program Core pick 25 June 2026
David Pariag (Meta)
- Why: The first real deployments of AI agents that touch production infrastructure — not write code, but operate systems (debug GPU failures, remediate outages). The access-control and guardrail problem this creates is one any operator of a large fleet will face within a couple of years; Meta is furthest along publicly.
- What: Meta's program-level learnings from enabling agentic access to its systems: building internal agents, deploying third-party agents, observing how agents interact with their infrastructure, and deploying guardrails to prevent destructive actions. A journey/lessons talk rather than a single system. (from abstract only)
- How: Not detailed in the abstract; expect case studies of what agents did to infra when let loose, and the containment mechanisms built in response.
- Where applicable: Meta-specific scale and internal tooling, but the problem space — authenticated agent identity, action scoping, blast-radius limits, observability of agent behavior — is universal for anyone running cloud or Kubernetes estates.