MLOps London November Talks: Deploying Production-Ready, Data-Centric Agents

AI agents are increasingly prevalent, but transitioning them from proof-of-concept to production-ready systems remains challenging, as highlighted in a talk by Seldon's Paul Bridi and Antanina Vertsinskaya. They discuss the fundamental characteristics of agentic systems, such as goals, tools, reasoning, and memory, and address major obstacles in production, including design, observability, and scalable deployment. The session features a live Kubernetes demo showcasing a marketing AI agent using Seldon Core and the LLM Module, emphasizing real-time tool-calling, memory capabilities, and an inference pipeline for handling substantial workloads. The presentation also explores the role of frameworks like LangChain, the benefits of modular data-centric inference graphs with Core 2, and the use of observability tools such as Kafka UI and Grafana/Prometheus. Upcoming developments include Python-first agent building, enhanced user interfaces, and improved support for reusable guardrails, offering guidance for those seeking to make AI agents resilient and reliable in real-world applications.