Agentic Practical Best Practices - For Real Engineers

Building AI agents for real engineering tasks? In this presentation, "Agentic Practical Best Practices," we bypass the hype and dive deep into the architecture of autonomous agents, context management, and complex AI workflows - built by and for real software engineers. Join me as I break down my battle-tested, custom agentic setup. Discover why you might want to skip off-the-shelf plugins and MCPs in favor of building your own CLI-based agent workflows. This talk covers everything from the core fundamentals of LLMs and context degradation to advanced engineering concepts like self-improvement loops, auto-allow mechanisms, sub-agent orchestration, and deploying a full "Review Fleet" for code and design reviews. You're welcome to follow me on my: Github: https://github.com/iholder101 Linked-In: https://www.linkedin.com/in/itamar-holder-39095b108/ Email: [email protected] Website: https://iholder.net/ Video Chapters: 00:00 - Introduction 03:13 - Part 1: The fundamentals 03:25 - What an LLM? 05:08 - What's an Agent? 07:24 - Why is context so important? 11:14 - Configuration layout overview 12:01 - Rules 14:34 - Skills 15:37 - Sub-agents 17:32 - Combining all the above 20:06 - Hooks 23:01 - Example: Better code reviewer 26:38 - Plugins? 30:21 - MCP vs. CLI 34:11 - Part 2: General best practices 34:22 - These tools demand to spend time 38:14 - Process is everything 39:58 - Aim for the future 41:36 - Autonomy 43:22 - Autonomy: Self-improvement 45:19 - Autonomy: Safe Environment 48:05 - Discipline 48:36 - Discipline: Evidence over assertion 52:05 - Blocking bad built-in features 53:37 - Part 3: A glance to my setup 55:45 - The "why" skill 58:14 - Session continuity 1:04:32 - Self-improvement Infrastructure 1:04:54 - The /improve skill 1:07:36 - Auto-memory & Awareness 1:09:54 - The Auto-Allow mechanism 1:19:03 - Integrating different LLMs 1:22:02 - Context Management 1:30:31 - Workflows 1:31:07 - Questions & Explorations 1:36:18 - Planning 1:41:00 - Code Review 1:49:28 - Design / VEP Review 1:52:39 - Implementation 2:04:36 - Vision for KubeVirt's AI-SDLC 2:07:30 - Final words