Understanding The Operational And Organizational Challenges Of Agentic AI
Summary<br />In this episode of the AI Engineering podcast Julian LaNeve, CTO of Astronomer, talks about transitioning from simple LLM applications to more complex agentic AI systems. Julian shares insights into the challenges and considerations of this evolution, emphasizing the importance of starting with simpler applications to build operational knowledge and intuition. He discusses the parallels between microservices and agentic AI, highlighting the need for careful orchestration and observability to manage complexity and ensure reliability, and explores the technical requirements for deploying AI systems, including data infrastructure, orchestration tools like Apache Airflow, and understanding the probabilistic nature of AI models.<br /><br /><br />Announcements<br /><ul><li>Hello and welcome to the AI Engineering Podcast, your guide to the fast-moving world of building scalable and maintainable AI systems</li><li>Seamless data integration into AI applications often falls short, leading many to adopt RAG methods, which come with high costs, complexity, and limited scalability. Cognee offers a better solution with its open-source semantic memory engine that automates data ingestion and storage, creating dynamic knowledge graphs from your data. Cognee enables AI agents to understand the meaning of your data, resulting in accurate responses at a lower cost. Take full control of your data in LLM apps without unnecessary overhead. Visit <a href="https://www.aiengineeringpodcast.com/cognee" target="_blank">aiengineeringpodcast.com/cognee</a> to learn more and elevate your AI apps and agents.</li><li>Your host is Tobias Macey and today I'm interviewing Julian LaNeve about how to avoid putting the cart before the horse with AI applications. When do you move from "simple" LLM apps to agentic AI and what's the path to get there?</li></ul>Interview<br /><ul><li>Introduction</li><li>How did you get involved in machine learning?</li><li>How do you technically distinguish "agentic AI" (e.g., involving planning, tool use, memory) from "simpler LLM workflows" (e.g., stateless transformations, RAG)? What are the key differences in operational complexity and potential failure modes?</li><li>What specific technical challenges (e.g., state management, observability, non-determinism, prompt fragility, cost explosion) are often underestimated when teams jump directly into building stateful, autonomous agents?</li><li>What are the pre-requisites from a data and infrastructure perspective before going to production with agentic applications?<ul><li>How does that differ from the chat-based systems that companies might be experimenting with?</li></ul></li><li>Technically, where do you most often see ambitious agent projects break down during development or early deployment?</li><li>Beyond generic data quality, what specific data engineering practices become critical when building reliable LLM applications? (e.g., Designing data pipelines for efficient RAG chunking/embedding, versioning prompts alongside data, caching strategies for LLM calls, managing vector database ETL).</li><li>From an implementation complexity standpoint, what characterizes tasks well-suited for initial LLM workflow adoption versus those genuinely requiring agentic capabilities?<ul><li>Can you share examples (anonymized if necessary) highlighting how organizations successfully engineered these simpler LLM workflows? What specific technical designs, tooling choices, or MLOps practices were key to their reliability and scalability?</li></ul></li><li>What are some hard-won technical or operational lessons from deploying and scaling LLM workflows in production environments? Any surprising performance bottlenecks, cost issues, or monitoring challenges engineers should anticipate?</li><li>What technical maturity signals (e.g., robust CI/CD for ML, established monitoring/alerting for pipelines, automated evaluation frameworks, cost tracking mechanisms) suggest an engineering team might be ready to tackle the challenges of building and operating agentic systems?</li><li>How does the technical stack and engineering process need to evolve when moving from orchestrated LLM workflows towards more complex agents involving memory, planning, and dynamic tool use? What new components and failure modes must be engineered for?</li><li>How do you foresee orchestration platforms evolving to better serve the needs of AI engineers building LLM apps? </li><li>What are the most interesting, innovative, or unexpected ways that you have seen organizations build toward advanced AI use cases?</li><li>What are the most interesting, unexpected, or challenging lessons that you have learned while working on supporting AI services?</li><li>When is AI the wrong choice?</li><li>What is the single most critical piece of engineering advice you would give to fellow AI engineers who are tasked with integrating LLMs into production systems right now?</li></ul>Contact Info<br /><ul><li><a href="https://www.linkedin.com/in/julianlaneve" target="_blank">LinkedIn</a></li><li><a href="https://github.com/jlaneve" target="_blank">GitHub</a></li></ul>Parting Question<br /><ul><li>From your perspective, what are the biggest gaps in tooling, technology, or training for AI systems today?</li></ul>Links<br /><ul><li><a href="https://www.astronomer.io/" target="_blank">Astronomer</a></li><li><a href="https://airflow.apache.org/" target="_blank">Airflow</a></li><li><a href="https://www.anthropic.com/" target="_blank">Anthropic</a></li><li><a href="https://www.anthropic.com/engineering/building-effective-agents" target="_blank">Building Effective Agents</a> post from Anthropic</li><li><a href="https://www.astronomer.io/airflow/3-0/" target="_blank">Airflow 3.0</a></li><li><a href="https://microservices.io/" target="_blank">Microservices</a></li><li><a href="https://github.com/pydantic/pydantic-ai" target="_blank">Pydantic AI</a></li><li><a href="https://www.langchain.com/" target="_blank">Langchain</a></li><li><a href="https://www.llamaindex.ai/" target="_blank">LlamaIndex</a></li><li><a href="https://leehanchung.github.io/blogs/2024/08/11/llm-as-a-judge/" target="_blank">LLM As A Judge</a></li><li><a href="https://www.swebench.com/" target="_blank">SWE (SoftWare Engineer) Bench</a></li><li><a href="https://www.cursor.com/" target="_blank">Cursor</a></li><li><a href="https://windsurf.com/editor" target="_blank">Windsurf</a></li><li><a href="https://opentelemetry.io/" target="_blank">OpenTelemetry</a></li><li><a href="https://en.wikipedia.org/wiki/Directed_acyclic_graph" target="_blank">DAG == Directed Acyclic Graph</a></li><li><a href="https://en.wikipedia.org/wiki/Halting_problem" target="_blank">Halting Problem</a></li><li><a href="https://arxiv.org/html/2410.15665v1" target="_blank">AI Long Term Memory</a></li></ul>The intro and outro music is from <a href="https://freemusicarchive.org/music/The_Freak_Fandango_Orchestra/Tales_Of_A_Dead_Fish/Hitmans_Lovesong/" target="_blank">Hitman's Lovesong feat. Paola Graziano</a> by <a href="http://freemusicarchive.org/music/The_Freak_Fandango_Orchestra/" target="_blank">The Freak Fandango Orchestra</a>/<a href="https://creativecommons.org/licenses/by-sa/3.0/" target="_blank">CC BY-SA 3.0</a>
