Why Vibe Coding Fails in Production| Krishna Kumar Sharma | Ex-Amazon AI Head | Omokai EP 63

AI agents can build a demo in a day. But what happens when they touch a production system with years of technical debt, undocumented decisions, security requirements, and real customers?In this episode of So What About AI Agents, Philippe Trounev sits down with Krishna Kumar Sharma, former Head of Engineering for AI at Amazon and founder of Omokai, to talk about what agentic software development looks like outside of greenfield demos and AI hype.Krishna introduces his D3 framework — Discover, Define, Deliver — an approach to AI-assisted engineering based on the same principles used by mature software teams: understand the system, define the work, execute deliberately, and review everything.We get into:• Why greenfield AI coding demos don't represent enterprise software development• How AI-generated technical debt can compound at enormous speed• Why spawning 20, 50, or 100 agents usually isn't the answer• “Token maxing” versus ROI maxing• Using different AI models to review and challenge each other's work• Why cheaper and local models can often handle implementation after good planning• Claude, Codex, Gemini, GLM and local/edge models• Prompt caching and whether context-optimization tools actually save money• Security risks created by executives and teams vibe coding directly into production• Why human review still matters in agentic engineering• The D3 framework for AI-assisted brownfield development• Why boring, structured engineering practices become even more important with AIIn the second half, we move from software agents into the physical world.Krishna explains how Omokai is developing voice-driven command-and-control systems for robots and drones, including autonomous systems capable of operating with AI at the edge.We discuss:• Voice-controlled robots and drone swarms• Running small language models directly on robotic systems• Human-in-the-loop controls for safety-critical actions• Guardrails for autonomous machines• Robotics interfaces such as ROS2, MAVLink and PX4• Operating robots without continuous cloud connectivity• Sensor fusion, LiDAR, vision and GPS-independent navigation• Defense, security, inspection, disaster response and caregiving applications• What happens when AI agents move from software into the physical worldThe central argument of the conversation is simple:More agents aren't automatically better. More tokens aren't automatically better. The goal should be producing more value for every dollar, model call, and engineering hour you spend.Subscribe to So What About AI Agents for conversations with founders, researchers, engineers and operators actually building and deploying AI agents in the real world.https://www.docsie.io