Agents went from demo to production faster than the engineering practices around them could mature. The agent loop — plan, call tools, reflect, repeat — is the most visible part of an agent system, but it's a small fraction of what's needed to run one reliably and securely. Underneath the loop sits a stack of concerns that frameworks don't solve for you: portability across rapidly obsoleting models, durable state for long-running runs, deployment pipelines that ship daily, runtime sandboxes for code that agents generate and execute, and the observability you need to debug a system that keeps changing its mind. Each is a discipline in its own right, with conventions still being invented. In this track, we'll work through each layer with practitioners shipping agents in production today, drawing out the lessons that turn a compelling demo into a system you can stake a business on.
Architecting for Agents: Beneath the Loop
Track Host
Julie Amundson
Senior AI Infrastructure Consultant, ex-Googler, ex-Netflixer
Since 2018, Julie Amundson has been an AI infrastructure leader, across Netflix, Google Cloud and an independent consulting practice launched in 2026. Her work spans the full stack — from TPU and GPU orchestration at scale on Kubernetes up through end-to-end data science productivity in Metaflow. At Google Cloud, she was on the GKE AI Infrastructure team, working on the schedulers and accelerator primitives that enterprises use to run distributed training and inference. At Netflix, she contributed to Metaflow and the notebook platform that data scientists across the company used to take models from prototype to production. Earlier, Julie was a founder of Order of Magnitude Labs, building Active Inference agents; before that, she helped build Netflix's first-generation streaming service and launch it in over 50 countries between 2008 and 2012. The thread throughout her career has been getting research-stage systems to run reliably in production at scale.