Generative AI on Kubernetes

Chapter 9. Running Agentic Applications in Production

In Chapter 8 we explored architectural patterns for AI-driven applications and introduced agentic workflows at the conceptual level. Now we shift from architecture to the practical challenges of running these systems in production. Because the AI landscape in 2026 still evolves so rapidly, technical details can become obsolete within months. Rather than cataloging frameworks that may vanish, we concentrate on operational patterns that endure across tools and standards. Our goal is to equip you with guidance you can apply regardless of the framework you choose.

This chapter addresses three core challenges for running agentic applications on Kubernetes:

Security

Agents interact with external tools and data sources, often on behalf of users. You need robust identity management, authentication patterns, and authorization controls that preserve user context while allowing agents to operate autonomously.

Agent coordination

Multiagent systems require standardized communication protocols. Agents must discover each other’s capabilities, delegate tasks, and track progress across service boundaries.

State management

Unlike stateless REST APIs, agents maintain conversational context across multiple turns. Production deployments require persistent storage patterns that survive pod restarts and support horizontal scaling.

This chapter covers two protocols that emerged as de facto standards in late 2024 when it comes to agent communication. ...