"Just Ask Codex": Inside OpenAI's Own Working Culture | Romain Huet, OpenAI
Episode von AI to the DNA von Christoph Magnussen · 5. Juni 2026 · 56 min
What happens when the best engineers at OpenAI stop writing code by hand? How does a coding tool turn into a command center for agents that work for days without you in the loop? And why has Germany quietly become one of OpenAI's top five markets for Codex?
In this episode of AI to the DNA, the podcast for everyone who wants to really understand and apply AI, I sit down with Romain Huet, Head of Developer Experience at OpenAI and one of the key people behind Codex. We meet in Paris at OpenAI's local office, right after the team shipped GPT 5.5 and a wave of new Codex capabilities. Romain previously led product at Stripe and is now building the layer where developers, and increasingly everyone else, meet the models.
You can also watch this episode as a video podcast on my YouTube channel.
We talk about why Codex started as a coding tool, why it became something much bigger, and how OpenAI's own teams, from engineering to finance, now use it as their daily working surface. Romain explains why the jump from GPT 5.2 to 5.3 Codex was the real turning point, and what it means that even the best engineers no longer write code by hand.
From there we go deep into the architecture behind it: the open-source Codex harness, the layer that connects models, tools, files and your computer. We discuss what a harness actually is, why OpenAI made it public, and why prompt engineering as we knew it is essentially gone. We also dig into the harder questions, sandboxing, computer use, prompt injection, and how the new code review capabilities of GPT 5.5 are already preventing real production incidents at companies like Datadog.
Finally, Romain walks us through what excites him most right now: /goal, the command that lets Codex relentlessly pursue an ambitious goal across days of uninterrupted work. We close on a simple but uncomfortable insight: the models are so good now that the real bottleneck has become our own imagination.
I'm Christoph Magnussen and at AI to the DNA, I'm not just asking how to use AI. I’m asking why it works the way it does. This podcast is for everyone who wants to move beyond tool-tourism and master the real craft of AI: Learning to think, work, and build side by side with intelligent systems!
☀️ 🎓 Die Blackboat AI Summer School: Vier Wochen Online-Sprint mit Hands-on-Aufgaben, um vom Kl-Ausprobieren zu echten Pro-Skills zu kommen. Vom Tool-Tourist zu eigenen Agent-Systemen mit unserer AI Summer School!
Spare jetzt 250€ mit dem Gutscheincode: ATTD-SUMMER
-> Jetzt anmelden auf https://academy.blackboat.com
⭐️ Wenn dir AI to the DNA gefällt, freuen wir uns über eine 5-Sterne-Bewertung. Damit hilfst du uns enorm, den Podcast sichtbarer zu machen und noch mehr Menschen mit unseren Themen zu erreichen. Danke für deinen Support!
(00:00) - Intro
(01:34) - Welcome to Paris: Tool Tourists vs. Champions League
(02:35) - The Codex Turning Point: From 5.2 to Days of Uninterrupted Work
(05:53) - Codex Without the Word "Code": Explaining It to Mittelstand
(08:12) - Inside OpenAI: From Model Speed to "Just Ask Codex"
(12:50) - From Chatbot to Harness: What Has Actually Changed
(16:57) - Why OpenAI Open-Sourced the Codex Harness
(18:36) - The End of Prompt Engineering: Just Talk to Codex
(22:39) - Monorepos: How Google, OpenAI and Blackboat Organize Code
(25:23) - Slack, Agent Signatures and the FTE Question
(28:16) - Trust and Sandboxing: How Codex Asks for Permission
(32:22) - Computer Use: When the Agent Takes Over Your Mac
(34:27) - The Numbers: 8x Codex Growth, Germany at 10x
(36:48) - Beyond Code: Onboarding Codex for Non-Developers
(38:36) - ChatGPT vs. Codex: One Model, One Harness, Two Audiences
(43:05) - Security: From Prompt Injection to Daybreak
(49:22) - Sparkle of AGI: /goal – Codex's Most Ambitious Mode
(54:11) - Still Too Shy: The Final Take and Goodbye
Weitere Episoden von „AI to the DNA“
- Alle bauen mit KI – aber wer versteht noch, was entsteht? | Florian Baader, CEO meinGPT 25. September 2026
- Federated Learning: Wie Krankenhäuser gemeinsam ein KI-Modell trainieren, ohne Daten zu teilen | Daniel J. Beutel, Flower Labs 11. September 2026
- Von Google-Traffic zu MCP-Servern: Als Perplexity bei Statista anklopfte | Marc Berg, Statista 28. August 2026
- KI-Kennzeichnungspflicht: Was sich für Unternehmen jetzt ändert | AI to the DNA: Update 21. August 2026
- Betting Everything on AI Native: How Box Rewires Itself for the Agentic Era | Aaron Levie, Box 14. August 2026
- Wie Anthropic intern mit Claude arbeitet und wie MCP zum Industriestandard wurde | David Soria Parra, Anthropic 31. Juli 2026
- Die Hälfte des Internets ist KI-Content und keiner will ihn lesen | Dr. Paul Elvers, FUNKE 17. Juli 2026
- 5 Milliarden Bewertung, 30 Mitarbeiter: Die Wette auf sich selbst verbessernde KI | Dr. Richard Socher, Recursive & you.com 3. Juli 2026
- Warum Wohlstand unser Innovationsproblem ist: Deutschland zwischen Angst und Aufbruch | Dr. Tina Klüwer, AI NATION 19. Juni 2026
- Deutschlands zweite Chance: Warum Agentic AI aufholt, was die Digitalisierung verpasst hat | Uwe Peter, Cisco Deutschland 27. März 2026