Production Safeguards for Autonomous AI Agents: Runtime Enforcement, Privacy, and Observability
Sai Nagabhairava
Video
Abstract:
As AI systems become increasingly autonomous and interact with tools, data stores, and external services, safety and privacy can no longer depend only on model behavior. This talk will discuss practical systems approaches for safeguarding production AI agents, including propagating sensitive context across agent and tool boundaries, observing security-relevant interactions, enforcing policy at runtime, preventing tool-layer bypasses, and designing failure-containment mechanisms such as human-in-the-loop escalation and kill switches. I will also discuss some of the engineering tradeoffs involved in building safeguards that must operate reliably and efficiently at production scale.
The discussion will focus on general architectural patterns and production lessons and will not include confidential company-specific information.
Bio:
Sai Nagabhairava is a Staff Software Engineer at Meta with more than 15 years of experience building large-scale distributed systems. His current work focuses on safety and privacy infrastructure for autonomous AI systems, including AI-agent safeguards, privacy observability, sensitive-context propagation, and runtime enforcement.
LinkedIn: https://www.linkedin.com/in/sai-nagabhairava
