“Engineers are becoming sorcerers” | The future of software development with OpenAI’s Sherwin Wu

Inside our walls, nearly everyone codes with Codex and every pull request gets an AI pass, and the vibe is that engineers are now orchestrating swarms of agents like modern-day wizards. If you do not build for where the models are headed, you will get left behind, because today is the low-water mark for capability.

My guest is Sherwin Wu, who runs engineering for OpenAI’s API and Developer Platform and has a front-row seat to how everyone is building; how much of your team’s code is now written by AI?

Managers like me find it faster to let Codex draft code and then guide it, and most engineers start with AI-generated code and refine. Ninety five percent use Codex daily, all PRs get AI review, and people who lean in open far more PRs as trust grows.

With that much automation, what does the engineer’s job become over the next couple of years?

Engineers are shifting from typing every line to leading many parallel agent threads, steering, checking, and giving feedback. Think SICP’s wizard metaphor or the Sorcerer’s Apprentice: it feels magical and high leverage, but you still need an adult in the room to prevent brooms from flooding the place.

People feel stress when agents stall or go off track; how do you handle that?

We even have a team running a fully AI-authored codebase with no manual escape hatch, and most failures come down to missing context. The cure is encoding tribal knowledge into the repo with clear docs, comments, structure, and files the model can reliably use.

How are you keeping code reviews from becoming a bottleneck as PR volume rises?

AI reviews everything and slashes review time to minutes, and our CI and lint loops are auto-fixed and retried by Codex, so humans can skim instead of grind. I would not claim every line in production is AI-written, but the bulk starts there now.

What is changing for managers, and how do you lead in this world?

AI supercharges top performers, so I spend extra time unblocking them and sharing their playbooks across the org. With tools that summarize work, pull org context, and draft reviews, managers can handle larger teams, and my philosophy is the surgeon model: anticipate blockers and have the scalpel ready.

Beyond the solo-unicorn meme, what second-order effects are we underestimating?

Expect an explosion of small, vertical SaaS—tens of thousands of durable ten to fifty million dollar companies that supply hyper-specific tools, which in turn let tiny teams punch way above their weight. That dynamic could reshape venture math and make distribution and audience even more valuable.

You’ve seen many AI deployments end up underwater on ROI; what is going wrong, and what works?

Outside our tech bubble, users often lack know-how, and top-down mandates without bottoms-up champions flop. Create a cross-functional tiger team of technical-adjacent power users to pilot workflows, codify best practices, and spread excitement so adoption sticks.

You argue customer feedback can mislead in AI; why is that?

Models improve so quickly that hand-built scaffolding people ask for today gets absorbed or obsoleted tomorrow, whether it was vector-heavy retrieval or rigid agent frameworks. Balance requests with where capabilities are trending and design for the near-future model so you do not trap yourself on a local maximum.

What should builders expect from the API and models over the next year?

Agents will handle coherent multi-hour tasks, which changes how you design feedback and supervision, and native speech models will leap in quality and unlock big enterprise audio use cases. Most products today are tuned for minutes; plan for hours.

What opportunity are you most bullish on right now?

Business process automation is huge and underrated—repeatable, SOP-driven work integrated with enterprise data is everywhere, and AI can make it faster and more reliable.

Founders fear the platform will compete with them; how should they think about OpenAI’s role?

The market is vast; winners succeed by building what people love, not by dodging us, and we run a neutral ecosystem where the same models ship to the API. ChatGPT’s app platform broadens reach, and free and affordable tiers push benefits widely because our mission is to spread this technology to everyone.

For someone ready to build, what does your stack let them do?

Use the Responses API to drive long-running agents at a low level, step up to the Agents SDK to orchestrate tools and sub-agents, add AgentKit and widgets for a polished UI, and use Evals to measure and improve; mix layers as you like.

Parting advice for listeners who do not want to miss this wave?

The next few years will be unusually fun—engage now, learn the tools’ edges, and let them into your daily work. Ignore the firehose, pick a couple products, wire them to your data, and build momentum.

Lightning round: books, shows, products, motto, and a housing-price surprise?

Read Anti‑Memetics Division and recent China books like Breakneck; I’m enjoying the new Jujutsu Kaisen season; Ubiquiti gear has been a joy to set up; my motto is never feel sorry for yourself; and in housing, high-voltage lines, floor plan quality, and curb appeal matter more than most expect.

Where can people find you, and how can listeners be helpful?

I’m on X as @sherwinwu, and I love hearing what you are building—reach out and tell me how we can help.

Thanks for listening; subscribe on your favorite app and find past episodes at Lenny’sPodcast dot com.