Just Talk To It - the no-bs Way of Agentic Engineering
The author has shifted to using OpenAI's codex CLI as their primary tool for agentic engineering, running 3-8 parallel agents simultaneously in a terminal grid to generate ~100% of code for a large TypeScript React application with multiple platform targets. Rather than over-engineering prompts or agent configurations, the key strategy involves managing "blast radius"—carefully scoping changes to minimize file impacts, enabling atomic git commits and easy rollback—while maintaining the ability to stop agents mid-execution and course-correct without losing work. The author prefers gpt-5-codex on mid settings over Claude Code due to its more pragmatic execution style and superior file context awareness, emphasizing that practical iteration and rapid feedback loops outperform elaborate prompt engineering or infrastructure like worktrees and PRs.
Was this useful?