Agentic AI is surprisingly easy to demo. Operating it reliably is a very different problem.

In one production platform, what started as an AI workflow evolved into a distributed system involving multiple dependent agents, FastAPI services, Elasticsearch-backed data services, execution workers, model endpoints, and workloads running on Kubernetes.

At that point, the difficult questions were no longer just about prompts or models:

Where should state and persistence live?
What happens when one step in a multi-agent workflow fails?
How do you separate agent logic from execution and data layers?
How do you observe a request that crosses several services and agent calls?
What should Kubernetes manage, and what should remain an application concern?
How do you debug a workflow that is technically "running" but not behaving correctly?

This session walks through the architecture and engineering decisions behind operating a multi-agent system on Kubernetes, including service boundaries, failure handling, deployment patterns, and observability using Prometheus, Grafana, and application-level tracing.

Rather than focusing on a particular AI framework, the session treats agents for what they eventually become in production: distributed workloads that need the same reliability, visibility, and operational discipline as any other cloud-native system.

The architecture and examples will be anonymized and will not contain client-sensitive information.