DiscussionTechnical

OpenAI Codex lead on the new shape of product work | Andrew Ambrosino

Andrew Ambrosino, product lead for Codex at OpenAI, discusses how AI is fundamentally inverting the product development process—implementation is no longer expensive, making taste and curation the critical challenges. He explains that with 90% of OpenAI's employees using Codex, the company is learning how to build products in an AI-native world where prototypes proliferate and medium selection matters more than process adherence.

Summary

Andrew Ambrosino shares how artificial intelligence is reshaping product development at OpenAI, particularly through the Codex application. The core transformation he describes is an inversion of traditional product work: historically, implementation was expensive, so teams invested heavily in upfront research, design, and documentation to de-risk decisions before building. Now, with AI capable of rapidly generating code and features, implementation has become cheap, which paradoxically makes the process harder. Instead of one carefully planned feature, teams generate dozens of parallel explorations, shifting the valuable work from implementation to curation and taste—deciding what's good, what fits the system, how to frame ideas, and what deserves investment.

Ambrosino emphasizes that taste encompasses multiple dimensions: aesthetic design, systems thinking about how features fit together, understanding product direction and themes, and the deeper abstractions connecting code to design intent. He notes that AI models still struggle with design because grading design quality involves subjective human judgment, cultural context, and the need for novelty—elements harder to incorporate into training loops than evaluating whether code compiles correctly.

On the evolution from Codex as a CLI to a full desktop application, he explains the unexpected finding that when internal employees across finance, marketing, legal, and other non-technical functions began using Codex, they never abandoned it for specialized tools supposedly better suited to them. This revealed that the product shape itself—a conversational interface that can see and understand code—had broader applicability than initially assumed. Rather than fragmenting into separate apps for different personas, the team is building one extensible platform that can work deeply with different verticals.

Ambrosino discusses the challenge of planning in an AI-native environment: precise long-term roadmaps create false precision, so planning must remain appropriately vague for nine months out while keeping near-term plans detailed. He shares the example of the Codex app itself—the exact same product would have failed if released in November but succeeded in February 2024, purely because model capabilities improved in between.

On design process versus implementation, he pushes back against the "design is dead" narrative, arguing instead that the medium for communication matters tremendously. Sometimes a document is the right first artifact, sometimes a prototype, sometimes code. The danger is prototype anchoring—people get attached to a production-quality prototype meant for exploration and over-weight it in decision-making.

He describes real usage patterns, including an internal editor named Brent who began using Codex for video editing by asking if it could edit videos, leading Codex to build its own Premiere Pro extension to interact with the specialist tool. This exemplifies the model where Codex acts as a coordinator that can use or extend other software rather than trying to replicate all functionality internally.

Ambrosino also discusses his own workflow evolution: initially building the Codex app with Codex itself, then shifting to using it for product discovery and coordination as his role expanded. He sets up automations to aggregate information from 3,000 Slack channels into daily briefs, coaching the model iteratively rather than writing detailed instructions upfront.

On roles and specialization, he argues against the extreme position that all roles will collapse into "builders." While implementation tools have democratized and role boundaries have blurred productively, disciplines still have knowable best practices and skill components worth preserving. Not everyone should work on everything—depth and breadth matter, and the gatekeeping aspects of roles (memorizing syntax, using exact tools) are eroding in helpful ways, but the underlying disciplines remain valuable.

He shares stories of failure and iteration: his startup ventures in regulated spaces that felt like constant failure, multiple micro-failures in merging Codex with ChatGPT where internal feedback was harsh but ultimately constructive, and the original Codex web interface that didn't work but later provided valuable learnings for subsequent iterations. He emphasizes that certain features built too early for current model capabilities aren't failures—they're artifacts to test against future models.

About this episode

<p><strong>Andrew Ambrosino</strong> leads development of the Codex desktop app at OpenAI. Nearly 100% of OpenAI employees—not just engineers—now use Codex weekly. A lifelong builder with a background spanning engineering, design, product management, and founding companies, he is now responsible for turning the Codex desktop experience into what he calls “the best desktop app that has ever existed, full stop.”</p><p></p><p><strong>In our in-depth conversation, we discuss:</strong></p><p>1. Why AI has completely flipped the product development process</p><p>2. What “taste” really means as a professional skill, and why it is emerging as the most valuable capability in an AI-first workplace</p><p>3. Why Andrew believes the Codex app would have failed if they launched it last November (vs. in February)</p><p>4. The “zone defense” model for how product managers at OpenAI operate when everyone can build anything</p><p>5. How roles are collapsed on Andrew’s team, and why eliminating the concept of roles entirely is a big mistake</p><p>6. How Andrew uses Codex to run his own workflows</p><p>7. The vision for a home base that coordinates work across ChatGPT, Codex, and the tools people already use.</p><p>—</p><p><strong>Brought to you by:</strong></p><p><a href="https://workos.com/lenny" target="_blank"><strong>WorkOS</strong></a>—Make your app enterprise-ready, with SSO, SCIM, RBAC, and more</p><p><a href="https://mercury.com/command?utm_source=lennys&#38;utm_medium=sponsored_newsletter&#38;utm_campaign=26q3_brand_campaign" target="_blank"><strong>Mercury</strong></a>—Radically different banking, now with Command</p><p>—</p><p><strong>Episode transcript:</strong> <a href="https://www.lennysnewsletter.com/p/openais-codex-lead-on-the-new-shape" target="_blank">https://www.lennysnewsletter.com/p/openais-codex-lead-on-the-new-shape</a></p><p>—</p><p><strong>Archive of all Lenny's Podcast transcripts:</strong> <a href="https://www.dropbox.com/scl/fo/yxi4s2w998p1gvtpu4193/AMdNPR8AOw0lMklwtnC0TrQ?rlkey=j06x0nipoti519e0xgm23zsn9&#38;st=ahz0fj11&#38;dl=0" target="_blank">https://www.dropbox.com/scl/fo/yxi4s2w998p1gvtpu4193/AMdNPR8AOw0lMklwtnC0TrQ?rlkey=j06x0nipoti519e0xgm23zsn9&amp;st=ahz0fj11&amp;dl=0</a></p><p>—</p><p><strong>Where to find Andrew Ambrosino:</strong></p><p>• X: <a href="https://x.com/ajambrosino" target="_blank">https://x.com/ajambrosino</a></p><p>• LinkedIn: <a href="https://www.linkedin.com/in/ajambrosino" target="_blank">https://www.linkedin.com/in/ajambrosino</a></p><p>• Website: <a href="https://ambrosino.io" target="_blank">https://ambrosino.io</a></p><p>—</p><p><strong>Where to find Lenny:</strong></p><p>• Newsletter: <a href="https://www.lennysnewsletter.com" target="_blank">https://www.lennysnewsletter.com</a></p><p>• X: <a href="https://twitter.com/lennysan" target="_blank">https://twitter.com/lennysan</a></p><p>• LinkedIn: <a href="https://www.linkedin.com/in/lennyrachitsky/" target="_blank">https://www.linkedin.com/in/lennyrachitsky/</a></p><p>—</p><p><strong>In this episode, we cover:</strong></p><p>(00:00) Introduction to Andrew Ambrosino</p><p>(02:30) How AI is changing the shape of product work</p><p>(06:32) When to use documents vs. prototypes</p><p>(10:25) What “taste” actually means</p><p>(12:06) Why AI is still bad at design</p><p>(16:18) Is the design process really dead?</p><p>(21:35) What the design process looks like on the Codex team</p><p>(23:41) Are product functions disappearing?</p><p>(27:22) Team structure</p><p>(30:12) IC vs. management</p><p>(31:37) Planning roadmaps</p><p>(35:16) Building features that don’t work yet</p><p>(38:13) The ambition problem: when you’re too AGI-pilled</p><p>(39:17) The latest frontier: loops and autonomous development</p><p>(52:05) How Andrew uses Codex to automate his entire job</p><p>(46:52) The power of computer use and browser automation</p><p>(49:10) Will we run all our SaaS apps inside Codex?</p><p>(52:05) The future vision for Codex</p><p>(57:20) The videographer who built a Premiere Pro extension with Codex</p><p>(59:30) Failure corner</p><p>(1:01:50) Lightning round</p><p>(1:07:03) BTS: How our producer uses Codex for editing</p><p>—</p><p><strong>References: </strong> <a href="https://www.lennysnewsletter.com/p/openais-codex-lead-on-the-new-shape" target="_blank">https://www.lennysnewsletter.com/p/openais-codex-lead-on-the-new-shape</a></p><p>—</p><p>Production and marketing by <a href="https://penname.co/" target="_blank">https://penname.co/</a>. For inquiries about sponsoring the podcast, email <a href="mailto:[email protected]" target="_blank">[email protected]</a>.</p><p>—</p><p><em>Lenny may be an investor in the companies discussed.</em></p> <br /><br />To hear more, visit <a href="https://www.lennysnewsletter.com?utm_medium=podcast&#38;utm_campaign=show-notes-no-free-preview-language">www.lennysnewsletter.com</a>

Key Insights

  • Ambrosino observes that because implementation is now cheap via AI, the expensive part of product work has shifted to taste, curation, and deciding what's worth building among dozens of parallel explorations
  • He argues that AI models struggle with design because the feedback mechanism requires human judgment about taste and culture, which is harder to systematize in training than evaluating whether code compiles correctly
  • The Codex app was originally designed as a developer tool, but when non-technical employees began using it instead of specialized tools ostensibly built for them, it revealed the product had broader applicability than anticipated
  • Ambrosino claims that in AI-native environments, long-term planning beyond three months must stay vague to avoid false precision, as model capability improvements can completely change what's viable
  • He contends that the exact same Codex product would have failed in November 2023 but succeeded in February 2024 due solely to underlying model improvements, not product changes
  • Ambrosino argues against viewing the design process as entirely dead, instead saying the death is tied to specific tools and day-to-day practices, while the conceptual overlay of design stages remains crucial
  • He observes that when Brent, an internal video editor, asked Codex to edit videos, Codex autonomously built a Premiere Pro extension to interact with the specialist tool rather than attempting to replicate the functionality internally
  • Ambrosino believes taste encompasses aesthetic design, systems thinking about feature integration, directional alignment, and the abstractions connecting code to design intent across a codebase
  • He states that the primary challenge in eliminating rigid role boundaries is not that specializations disappear, but that gatekeeping aspects of roles (syntax memorization, specific tool mastery) become less relevant while underlying skill and discipline remain valuable
  • Ambrosino describes how he set up personal automations in Codex to aggregate information from 3,000 Slack channels into daily briefs, then iteratively coaches the model when it misses priorities rather than writing detailed initial instructions
  • He claims that features built before models are capable enough aren't failures but rather artifacts to test against future model capabilities, citing the evolution from Operator to Atlas to Codex as iterations of the same feature concept
  • Ambrosino argues that in product coordination at scale, zone defense among product people prevents redundant coverage, requiring them to spread out to identify and fill gaps where teams are misaligned or exploring in conflict

Topics

AI-driven inversion of product development processImplementation becoming cheap, curation becoming valuableThe concept of taste in product workDesign challenges for AI modelsCodex product strategy and evolutionPlanning under AI capability uncertaintyRole collapse and specialization in AI-native teamsPrototyping and medium selection in product workDogfooding and personal usage patternsIntegration with specialist tools via extensions and APIsFailure and iteration in AI product developmentDesigning for multiple user personas with one platform

Transcript

90% of people at OpenAI use Codex. Not 90% of engineers, that's 90% of the entire company. You had this tweet the other day where you said that you intend to make Codex the best desktop app that has ever existed. Yeah. The quality bar for Codex had to be so high that there was never a hesitation that you have opening this app to do the next thing. That this was your natural choice, just like people have come to open a browser tab. That's true, I know. There's numbers constantly coming out about the records you guys are setting for usage. I don't know, we'll see. A lot of people seem to like the app. Why do you…

Full transcript available for MurmurCast members

Sign Up to Access

More from Lenny's Podcast: Product | Career | Growth

Get AI summaries like this delivered to your inbox daily

Get AI summaries delivered to your inbox

MurmurCast summarizes your YouTube channels, podcasts, and newsletters into one daily email digest.