Building AI That Doesn’t Break

Argo Workflows is a powerful workflow orchestration engine, built for Kubernetes. But like most systems built on Kubernetes, you usually use YAML to define resources. In Argo Workflows, these Workflow resources include complex DAGs (Directed Acyclic Graphs), with dependencies and parameters being passed between tasks, and artifacts being uploaded to storage. If you’re trying to express this imperative workflow logic using YAML, you’re in for a rough ride – a single Workflow can eventually span thousands of lines, leaving you lost in a sea of indentation and key-value pairs whenever you want to make changes.

Enter Hera, the Python SDK for Argo Workflows. By treating Python functions as the unit of work that run in separate pods, orchestrating them becomes as effortless as calling the functions to create a DAG. With Hera, you get reusability and unit testing for free, something not possible in YAML, and you don’t need to use separate languages for your business logic and orchestration logic, keeping everything under one roof.

If you're tired of wrestling with YAML but want to run your Python jobs on Kubernetes, this session is for you!

Building AI applications as workflows is not new: many popular frameworks let you define DAGs to interact with LLMs and manage data pipelines. But real-world applications require more than just function chaining. Instead, they need careful orchestration to wait for human input (which might take hours or days), retry failed API calls, parallelize tasks, recover from server crashes, or prevent duplicate or missed updates.

In this talk, we'll introduce the MLOps community to durable workflows, a solution where each step is automatically checkpointed to a durable store, and workflows are guaranteed to run to completion. This helps AI applications automatically recover from interruptions, restarts, and other failures without losing progress.

We'll walk through real examples of building robust AI pipelines and agents using durable workflows, showing how they can manage complex logic and external dependencies while maintaining strong correctness guarantees.

Function calling gives LLMs access to tools, but it breaks down in real-world, multi-step workflows. In this talk, Alan Nichol introduces Process Calling—a more reliable, stateful alternative where LLMs collaborate with long-running business processes. This approach reduces flakiness, improves modularity, and makes debugging easier. If you're building customer-facing AI, Process Calling is the pattern you need to move beyond “prompt and pray” and into production-ready assistants.