AI systems are rapidly evolving from passive models into autonomous agents capable of executing complex workflows across APIs, services, and cloud infrastructure. But deploying these systems in production introduces a new class of distributed systems challenges.

In this talk, we explore agentic infrastructure - the systems required to safely run AI agents at scale. Drawing from real-world experience building large-scale infrastructure at Meta, we will examine how agent runtimes interact with APIs, tools, and microservices, and why traditional reliability patterns break down.

We’ll cover:

1. Failure modes unique to AI agents (hallucinated actions, cascading retries)
2. Guardrail architectures and action validation pipelines
3. Observability for non-deterministic systems
4. Cost and safety control mechanisms

Attendees will leave with a blueprint for building production-grade AI systems that are reliable, observable, and safe.