Stenobird

Episode

Building Durable AI Agents

Podcast
Practical AI
Published
Jul 9, 2026
Duration seconds
2799
Processing state
processed
Canonical source
https://share.transistor.fm/s/facb92e2
Audio
https://pscrb.fm/rss/p/dts.podtrac.com/redirect.mp3/media.transistor.fm/facb92e2/33d1bc96.mp3
JSON
/v1/public/podcasts/practical-ai/episodes/building-durable-ai-agents
Markdown
/podcast/practical-ai/building-durable-ai-agents.md

Actions

  • POST https://stenobird.com/v1/public/podcasts/practical-ai/episodes/building-durable-ai-agents/transcription-requests
    Idempotently request low-priority transcript generation for this episode.
  • GET https://stenobird.com/podcast/practical-ai/building-durable-ai-agents.md
    Read the agent-friendly Markdown representation of this episode resource.

Summary

Moving AI agents from experimental demos to production requires applying MLOps principles to handle long-running, complex workflows. This episode explores the infrastructure needed for durability, state management, and scalable agent fleets.

Topics

  • AI Agents
  • MLOps
  • Agentic Workflows
  • Infrastructure
  • State Management
  • ZenML
  • LLM Orchestration
  • Software Engineering

Highlights

  • Main idea: Agentic workflows require a shift from simple request-response patterns to durable, long-running execution environments
  • Practical takeaway: Enterprises should invest in internal agent platforms to manage complex business processes and maintain control over infrastructure
  • Failure mode: Relying solely on model-specific features creates vendor lock-in and fragility when model availability or regulations change
  • Technical necessity: Robust systems must implement state management, retries, and observability to handle inevitable failures in multi-step agent loops
  • Future trend: As LLM performance commoditizes, the competitive advantage will shift toward the quality of the underlying agent harness and runtime

Chapters

  1. 1:00 Introduction to Kitaru: An introduction to Hamza Tahir and ZenML's new project, Kitaru, which focuses on making AI agents durable.
  2. 4:00 The Evolution of MLOps to AgentOps: Discussing how the principles of DevOps and MLOps are being reinvented to support agentic workflows and software engineering standards.
  3. 8:00 New Workload Challenges: Exploring the complexities of state management, retries, and replays in modern AI workloads.
  4. 11:00 Scaling Multi-Agent Architectures: The rise of agent fleets and swarms, and the infrastructure challenges that emerge at scale.
  5. 15:00 The History of AI Evaluation: Reflecting on the rapid evolution of evaluation methods from manual Python scripts to automated systems.
  6. 18:00 Decoupling Logic from Models: Why business value and infrastructure should remain independent of specific LLM providers to ensure business continuity.
  7. 22:00 Architecting for Durability: Understanding the different types of agent workloads and how to build resilient, non-fragile agent harnesses.