Summary
Peter Yang sits down with full-stack engineer Ras Mic, who teaches 100,000 people to code with AI, to unpack the agentic engineering system he uses to ship real apps. Rather than the “software factory” that fires off a task and hopes for the best, Mic describes a model- and harness-agnostic workflow built from one agents.md file and a handful of reusable skills.
The workflow centers on four steps. First, isolate: a “new feature” skill spins up a fresh git worktree off main so several coding agents can work in parallel without stepping on each other, as Mic demonstrates with multiple Codex agents running at once. Then come writing code to a specific standard, proving that a bug exists before fixing it, and an evidence-driven review loop using the Greptile code review agent, where each pull request gets a confidence score and feedback. Mic notes that first PRs from most models typically land around four out of five.
The conversation also covers choosing between terminal and desktop app, using browser and computer use for testing, and how Mic compares current frontier models for coding. The core message is that, even with powerful agents, the best results still come from keeping a human in the loop and testing each feature yourself.
📺 Source: Peter Yang · Published October 11, 2026
🏷️ Format: Interview
