Descriptions:
In this episode of the David Ondrej podcast, senior developer Mickey breaks down what he calls “agentic engineering” — a disciplined, structured approach to AI-assisted development that he argues has superseded the looser “vibe coding” era. With roughly 95% of his code now AI-generated, Mickey shares the specific stack and mental model behind his workflow.
The conversation covers the distinction between a model and a harness: tools like Cursor, Claude Code, and OpenAI Codex wrap models with file access, tool calls, system prompts, and agent loops that determine how effectively a model performs on real codebases. Mickey explains his current model preferences — GPT-5.5 for large, complex codebases and Opus 4.7 Max for UI and frontend work — and why harness choice matters as much as model choice. He also discusses why Svelte is a better fit for AI-generated code than React, citing its closer alignment to standard HTML and TypeScript syntax that models handle reliably.
A recurring theme is the importance of human judgment within agentic loops: defining clear end states, providing structured context, and knowing when to intervene. Mickey frames this as the real skill differentiator in 2026 — not prompt writing, but engineering the feedback loop itself. Practical and opinionated, this episode is a useful reference for developers looking to move beyond ad-hoc AI assistance toward reproducible, high-output agentic workflows.
📺 Source: David Ondrej · Published May 17, 2026
🏷️ Format: Podcast
