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

Today’s a rare in‑person episode. My guest is Andrew Emversino, who leads product and engineering for OpenAI’s Codex app. Codex has become a go‑to for building software and everyday work—organizing files, writing docs, crunching data, even triaging email. Since January, usage has grown about six times, with over five million weekly users, and inside OpenAI nearly everyone relies on it weekly. Stick around to the very end for a quick post‑recording clip with our producer on how he uses Codex in editing. Andrew, welcome—let’s dig into how AI is reshaping product work.

Ninety percent of OpenAI reportedly uses Codex. You also said you want to make it the best desktop app, period. What’s driving that bar?

I want the quality so high there’s no pause before opening it—like reaching for a browser tab. Tons of people seem to love it, but the ambition is what keeps us honest.

Why are current AI systems still weak at design, and do you expect that to change?

Design is hard to grade because human taste is part of the feedback loop, and research has focused more on capabilities that speed up AI progress, like code correctness. Models will get better, but culture, novelty, and the deep link between visual systems and underlying abstractions are tricky; aligning semantics across the codebase and UI still feels out of reach right now.

How has the shape of product teams changed compared to a few years ago? Are roles collapsing?

Implementation got cheap, so everyone can spin up working prototypes fast. The scarce part is selection and framing—taste. I’m wary of declaring roles dead; the crafts still matter. I like dissolving hard boundaries, not erasing disciplines.

Docs used to precede builds. Now people jump to prototypes. When is that the wrong first move?

Pick the medium that fits the question. If you need clarity in a fuzzy area, write. If you need to stress‑test an interaction, build. Over‑anchoring on a shiny proto that only meant to explore is dangerous. The bigger skill is knowing what to work on and how to communicate it.

Given your cross‑functional background, how do you run design for Codex?

We have overlap by design. Designers, PMs, and engineers all code, and your role is basically the average of how you spend time. We dogfood relentlessly, even when it slows us down, so the product improves to match the way we actually work.

Some teams want to call everyone a builder. Does that future make sense?

Don’t throw out the playbooks. Kill the lane‑policing, keep the standards. It’s easier to switch roles now because tools are less of a gate, but each craft still has real best practices.

What’s the Codex team footprint, and who thrives there?

The core team is in the dozens, but it’s the tip of a much larger effort across OpenAI. We hire high‑agency people with strong taste—many former founders—and keep teams IC‑heavy with big scopes.

Planning must be wild with fast model shifts. How do you approach roadmaps?

Detail the near term; keep the long term loose or you’re pretending. Prototype a wide set, then revisit as models improve. The same product can flop in November and click in February purely from model gains.

How do you balance ambition with the grind of polish?

You need both streams. Bottom‑up spikes can disrupt the mainline, while others harden reliability. We design the org so those cycles can coexist.

With agents and loops, what do cutting‑edge teams look like now?

Most code is model‑written; the question is how supervised it is. We’re pushing on refactors, deletion, abstractions, and choosing the right features, but we’re not at “set a goal and let it run the company” yet.

How do you personally use Codex day to day as a product and engineering lead?

I built the app with the app, then shifted it to discovery and release orchestration. It assembles briefs from Slack and repos, pings me with questions, and I steer it with quick prompts to adjust focus.

If someone wants that setup, what’s the workflow?

Create a scheduled automation, connect Slack, describe what to watch, and coach it after a few runs. If a connector’s missing, the app will offer to add it, or it can fall back to browsing or even computer control.

I watched it navigate the Google Cloud Console to wire up mail filtering. Wild. How do you decide between browser, extension, and full computer use?

We’re tuning those boundaries so it picks the fastest, safest path—connector, in‑app browser, Chrome via extension, or direct control—based on what the task needs.

Do you see Codex becoming the place where we use our SaaS apps with an agent beside us?

We think of it as your home base. Sometimes you do the whole task inside; other times it opens and drives Excel, Notion, or a browser, then brings you back. It should start, coordinate, and automate your work across tools.

How are Codex and ChatGPT coming together in that vision?

People wouldn’t leave Codex even for role‑specific apps, so we’re unifying strengths. Build general primitives that can go deep for legal, finance, research, and more, while keeping entry simple.

Share that Premiere Pro example—it captures what’s possible.

Our videographer had Codex edit project files and, when it hit limits, Codex wrote a Premiere extension it could command. That’s the pattern: integrate with expert tools, then extend them.

Fail Corner: a miss that taught you something vital?

I spent years at startups that struggled or sold for parts. Inside OpenAI, we still ship things that get shredded in long threads, then we iterate. Timing plus persistence matters more than a perfect first try.

Lightning round: recs, shows, products, motto, and which role is toughest?

I’m deep in kids’ books—The Gruffalo, and the wonderfully subversive The Big Orange Splat. The new Magic School Bus is a hit at home. I admire Linear’s craft. Roles are blending, but each still takes real skill.

That’s a wrap. Andrew, thanks for coming by. If you enjoyed this, subscribe on Apple Podcasts or Spotify, drop a rating or review, and find past episodes at Lennyspodcast.com.

After hours thought: don’t wed yourself to a fixed process. Stay married to outcomes you’re uniquely good at, and be willing to change how you work. Thriving with AI demands unusual self‑awareness, and constant relearning isn’t for everyone.