Just Talk To It - the no-bs Way of Agentic Engineering
This title could be clearer and more informative.Try out Clickbait Shieldfor free (5 uses left this month).
A developer shares their evolved workflow for AI-assisted coding, focusing on practical approaches over complex tooling. They use multiple parallel instances of OpenAI's codex CLI as their primary development environment, running 3-8 agents simultaneously in the same codebase. The author advocates for simplicity: direct conversation with AI agents, minimal custom tooling, and avoiding over-engineered solutions like elaborate subagent systems or RAG implementations. Key practices include atomic git commits by agents, using screenshots for context, spending 20% of time on AI-driven refactoring, and developing intuition about task complexity ("blast radius"). They prefer GPT-5-codex over Claude Code for its language style, efficiency, and reliability, despite some quirks. The workflow emphasizes speed and pragmatism, with agents handling virtually all code writing while the developer focuses on architecture, design decisions, and feature planning.
Table of contents
Context & Tech-Stack #Harness & General Approach #Model Picker #How I write prompts #Web-Based Agents #The Agentic Journey #But Claude Code can Background Tasks! #What about MCPs #But the code is slop! #Do you do spec-driven development? #Show me your slash commands! #What other tricks do you have? #How does your Agents/Claude file look like? #So GPT-5-Codex is perfect? #Conclusion #Sort: