Feb 21, 2026 · 1h 3m · 20vc
OpenAI's Codex Lead: Why Coding as We Know It is Over · 20VC with Harry Stebbings
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
OpenAI's Codex Product Lead, Alexander Emberikos, discusses the future of software engineering, explaining how AI is shifting developers from manual coding to task delegation and automated code reviews. He outlines OpenAI's strategic vision for user interfaces, agentic web tools, open standards, and the shifting dynamics of startup defensibility in the AI era.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Harry holds 23.2% of the talking time here. How this is scored →
Harry as informed peer 4.8 Guest teaching 5.4 Guest disagreement 2.6 Harry pushing back 3.8
speaking balance: gold is Harry, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 9:03 Direct Rejection of Enterprise FDE Premise
Alexander explicitly rejects the premise established by a previous guest on Harry's podcast, stating 'I actually disagree with that entirely' regarding the necessity of FDEs for enterprise AI adoption.
Hardest push from Harry ▶ 30:41 Challenging WAU vs DAU MetricsHarry forcefully refuses Alexander's framing of Weekly Active Users as a North Star metric, arguing that if Codex is replacing the IDE, it must be evaluated on Daily Active Users.
Biggest teaching moment ▶ 2:24 Historical Paradigm of Automation Increasing DemandAlexander educates Harry on historical technological shifts—from assembly code to high-level languages and Bletchley Park tabulations—demonstrating how automating specific tasks dramatically increases overall demand for engineers.
Harry holds his own ▶ 47:24 Detailed Nuance on SaaS Category SurvivalHarry demonstrates deep market knowledge by resisting broad claims of SaaS death, breaking down specific software dynamics across Monday.com, ServiceNow, Salesforce, and customer support platforms.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Harry as informed peer | Guest teaching | Guest disagreement | Harry pushing back | Why |
|---|---|---|---|---|---|---|
| Motivation: Winning vs. Avoiding Losing | 3 | 6 | 2 | 3 | Harry introduces Elon Musk's premise that coding will be automated, but Alexander reframes the question using historical examples like assembly code and Bletchley Park punch cards to show how automation increases total engineering demand. Alexander also notes that the talent stack is compressing toward full-stack roles, rendering traditional PMs largely redundant. | |
| Prompting Laziness & The Human Bottleneck | 3 | 5 | 2 | 2 | Alexander engages Harry in a Socratic exchange to highlight how human laziness and prompt friction act as the primary bottlenecks to AGI usage. Harry willingly participates, confirming his own usage patterns while Alexander explains that AI must become effortless and contextual rather than prompt-dependent. | |
| Productizing Prompts vs. Open-Ended Individual Tools | 5 | 6 | 4 | 4 | Harry cites previous guest Matt Fitzpatrick regarding enterprise FDE adoption, but Alexander directly rejects this thesis, stating he disagrees entirely. Alexander argues for open-ended, individual-focused tools rather than top-down, specialized enterprise integrations. | |
| Three Phases of Agents & Safe Enterprise Browsing with Atlas | 6 | 6 | 3 | 6 | Harry pushes back firmly on Alexander's individual-tool model, citing enterprise data security, permissioning, and employee capability limitations. Alexander defends his vision by outlining three agent phases and explaining how controlling the browser layer with Atlas enables safe, agentic enterprise browsing. | |
| Inference Speed, Competitive Pressures, and PLG vs. Sales/Marketing | 5 | 5 | 3 | 4 | Harry brings up the Cerebras partnership and quotes Jason Lemkin's thesis that inference spending will replace sales and marketing. Alexander disagrees with Lemkin's premise, arguing that customer acquisition and relationship building become even harder as market noise increases. | |
| Internal Codex Usage & The Ergonomics of Delegation | 4 | 5 | 2 | 3 | Harry asks for internal OpenAI metrics on Codex usage compared to competitors. Alexander explains how GPT-5.2 Codex shifted developer behavior from interactive pair-programming to full task delegation without opening IDEs. | |
| Autonomous Code Reviews & High-Signal Feedback | 5 | 5 | 2 | 3 | Harry quotes Tom Blomfield regarding low switching costs between AI coding providers. Alexander details how explicit plan modes prevent AI slop and how high-signal automated code reviews reduce false positive criticisms. | |
| Open Standards and the Mission to Distribute Intelligence | 5 | 7 | 4 | 6 | Harry references 'Seven Powers' to challenge OpenAI's retentive moat when open-sourcing standards like agents.md. Alexander reframes OpenAI's core mission around distributing intelligence—even serving competitors—which leaves Harry openly baffled from a VC perspective. | |
| Defining Winning: Compute Moats vs. Active User Metrics | 6 | 5 | 2 | 6 | Harry presses Alexander to choose between GTM, product execution, or compute as the winning factor. When Alexander reveals Weekly Active Users (WAU) as their primary internal North Star, Harry directly challenges whether WAU is frequent enough for an IDE replacement, forcing Alexander to concede DAU is better. | |
| The Future UI of AI: Chat versus Specialized GUIs | 5 | 5 | 2 | 3 | Harry cites Anish Acharya's counter-argument that chat interfaces are insufficient for most users who prefer graphical GUIs. Alexander agrees partially, framing conversational UI as the primary pillar supplemented by specialized graphical tools for power users. | |
| Agent-to-Agent Workflows & Acquiring Task Data | 5 | 6 | 2 | 3 | Harry asks about agent-to-agent workflows and asks if Anthropic holds a coding data moat. Alexander explains that coding data is plentiful, but non-public knowledge work task trajectories represent the true upcoming data bottleneck. | |
| Consumer Coding Democratization & Free ChatGPT Access | 4 | 5 | 3 | 2 | Harry asks if Codex competes directly with consumer platforms like Replit and Lovable. Alexander highlights Codex's expansion to free ChatGPT tiers, massive growth metrics, and makes a cheeky comment about competitor SOTA models lasting only 20 minutes. | |
| Benchmarks vs. Vibes & The Super Assistant Consolidation | 4 | 6 | 3 | 3 | Harry asks about terminal market composition across model providers. Alexander dismisses domain-specific agent fragmentation, drawing on his Dropbox history to argue that single super-assistants create unbeatable centers of gravity like Slack did for workplace communication. | |
| The Future of SaaS and the Distribution-First Founder | 7 | 5 | 3 | 6 | Harry pushes back against extreme claims that SaaS is dead, providing granular breakdowns of Monday.com, Salesforce, ServiceNow, and customer support tools. Alexander agrees that systems of record and human relationships survive, but advises investing in distribution-first founders. | |
| War for Talent & Selective Hiring of Product Fits | 4 | 5 | 1 | 2 | Harry asks about talent war dynamics and advice for computer science graduates. Alexander advises young engineers to demonstrate agency, taste, and high-quality public projects rather than traditional resumes. | |
| Quick Fire: Lessons from Dropbox & Inference Margin Pressures | 6 | 5 | 3 | 5 | Harry presses on unit economics and inference margin compression versus traditional software margins. Alexander argues that winning the immediate agent distribution race justifies taking short-term margin hits. | |
| Quick Fire Continued: AI Trajectories, Competitors, and Grandma's AI | 4 | 5 | 3 | 3 | In a rapid-fire sequence, Alexander reflects on the mistake of unlimited Codex Cloud pricing, predicts the end of manual code editing and CI deployment within five years, and shares his personal goal of making AI accessible to non-technical users like his grandmother via messaging apps. |
Statements from this episode (44)
Emberikos: OpenAI serves its trained models to direct competitors
“Actually, our job is the distribution of intelligence, and this is really unintuitive, but like, we put all this effort into training these models, and then we serve these models to our competitors.”
Emberikos: AI automation will create more software builders in five years
“Yeah. And I, you know, sometimes we change what terms mean, right? Like the term computer now refers to something else, but now we have the term software engineer. And so I definitely think we'll have many more builders.”
Emberikos: OpenAI Codex team is structured almost entirely as full-stack engineers
“Whereas like now, at least if I think about the Codex team, like there's very few, like that's much less the case and things are much more full stack.”
Emberikos: Codex users currently use the tool tens of times daily
“When we look at how often, ah, you know, codex users are using codex, it's like kind of this, like, tens of times kind of range.”
Emberikos: AI will eventually assist humans tens of thousands of times daily
“I think AI should be helping us tens of thousands of times per day. You know, compute budget permitting, well, and we'll get there over time.”
Emberikos: Future AI will eliminate prompting and proactively assist users
“And so I think the world we want to get to is one where to use AI, you don't really need to, like, figure out the right way to prompt. It's just super easy for you. And you don't even need to recognize that AI could help you. It's just, like, knows you, connec…”
Emberikos: Enterprise AI adoption does not require forward-deployed engineers
“So, so even though I am literally hiring FDs and if you're an FD, please apply for a job with me. I actually disagree with that entirely.”
Emberikos: AI tools outside coding should be open-ended, not vertically specialized
“Is not overly build it like, okay, this is AI capabilities, but only specifically for finance, only for specifically for this workflow, but actually build a much more open-ended tool that someone can just use for any given task creatively.”
Emberikos: Top-down enterprise AI deployment massively under-leverages AI potential
“What I've seen is that When we do these things top down, we end up, like, massively under leveraging the potential of AI in, like, helping that company.”
Emberikos: OpenAI is building Atlas browser for safe enterprise agentic browsing
“And, you know, I think it's quite unusual, like in OpenAI, we're building a browser, Atlas, right? And you might wonder why. And there are many reasons why, but I think one of the key reasons is that by building a browser, we can build sort of S and by control…”
Emberikos: AI inference market will not become a monopoly
“This is just my opinion, but I don't think we're going to end up in, like, this kind of monopolistic world. I think there is so much competitive pressure that there'll be, like, multiple answers to this.”
Emberikos: OpenAI recently accelerated API model serving by 40 percent
“So we recently rolled out a change where in the API, like, those models are served, like, 40% faster, and in Codex they're served, like, a quarter fast, 25% faster.”
Emberikos: AI will not eliminate enterprise sales and marketing teams
“But going back to the sales and marketing thing, like, I don't think that goes away because I think that's as, like I said, I think that's just gotten harder as the markets, any given market gets more competitive with more software out there.”
Emberikos: The vast majority of OpenAI code is now written by AI
“I don't have a percentage stat for you, but I would say, like, the vast majority of code is written by AI, and I would say that now, probably, like, most people are not even, like, opening IDEs. Maybe if they are opening IDEs to, like, maybe you want to own th…”
Emberikos: OpenAI intentionally omitted text editing features from the Codex app
“We explicitly didn't build editing into the Codex app because we wanted it to be really clear how you're meant to use it. So, you know, it has a lot of affordances for managing multiple agents, for delegating for reviewing changes. It has really prominent skil…”
Emberikos: Plan reviews are becoming more important than code reviews
“Review of the plan is actually something that's becoming more important because we're entering more of this like delegation phase of working with agents.”
Emberikos: Nearly all OpenAI code is automatically reviewed by Codex
“Nearly all code at OpenAI is reviewed by Codex automatically whenever you push it to a good repo.”
Emberikos: Nearly all AI agents except Claude adopted OpenAI's agents.md standard
“Pretty much every agent except Claude uses agents.md, which is awesome.”
Emberikos: AI agents gain enterprise stickiness by connecting to external systems
“As agents start to do work that is not writing code, but more general work, again, for software engineers or beyond for any builder, they're gonna need to start interfacing with other systems. Right? So as they start, maybe your agent is talking to Sentry, rig…”
Emberikos: Active users is OpenAI's primary success metric, not revenue
“It's actually not revenues, the primary is active users.”
Emberikos: AI's next phase this year is executing tasks, not retrieving information
“And I think the next phase that we'll see this year is like for any task I need to do, as opposed to just get information, I go to this text box or this input and something happens that helps me, even if it's not the full task, even if it's only a small part o…”
Emberikos: Future AI will pair conversational interfaces with bespoke power-user GUIs
“So I think what we're going to have is that we'll have chat or voice, basically conversational interface will be sort of the pillar of everything that you can talk to about anything. And then you can add into any group chat or whatever so it can, like, discove…”
Emberikos: Optimal software interfaces for AI agents mirror those for humans
“We've noticed as we build Codex that the best, like, the best interfaces for Codex to do work are also tend to be the best interfaces for humans.”
Emberikos: OpenAI has plenty of training data for AI coding models
“I think that from what we've seen, and you know, I would defer to my research team on this, but I feel like we feel like we have plenty enough data to build really good coding models.”
Emberikos: Knowledge work training data is harder to acquire than coding data
“Yeah, I think that, that kind of knowledge work task distribution is, like, much harder than coding.”
Emberikos: OpenAI uses external vendors for large data campaigns to move quickly
“Yeah, I mean, I think the way that we think about these things is just like, how do we move as quickly as possible? And so, you know, getting, becoming able to set these things up in-house is, like, very expensive in time, and we're a small team. So what I hav…”
Emberikos: Free ChatGPT users will use Codex over specialized coding tools
“And so I think we're definitely going to see people with like a free chat GPT plan coming in and just like building simple things where they otherwise might have gone to a specialized tool.”
Emberikos: OpenAI Codex grew 20x since August and doubled since December
“I feel like the public metric we have was like, since August, we grew by like, 20 X. And then like, even like late in the year, we like doubled from December to now.”
Emberikos: OpenAI Codex is pivoting back to building cloud-based autonomous agents
“The first is I actually want to get back to cloud. When we pivoted our strategy from, like, building the cloud, like, focusing on the cloud agent last year to working interactively, the thinking was very simple. It was just, and it's kind of like what I was te…”
Emberikos: Writing code is now trivial; code review is the main underinvested bottleneck
“Code gen writing code has become, like, you know, basically trivial now. But the hard part is, like, what you were talking about with, like, code review, right? Like, how do we know the code quality is good? How do we know we're doing the right things? And tho…”
Emberikos: OpenAI aims for AI agents that run micro-systems without human review
“So, like, I think we want to get to a world where you can have an agent that is un-bottlenecked, right? That you trust to, like, own an entire microsystem or internal tool or whatever, and can do the full iterative loop, including feedback from users, without …”
Emberikos: User evaluation of AI models is driven by vibes
“Like whenever I talk to any, like even internally or even talking to like customers of our models, I'm always surprised by how vibes based the evaluation of how it feels to work with the model is.”
Emberikos: Coding agents are the only domain with clear AI PMF
“So we're, so we only have PMF for coding agents, like, in the industry overall, I would say, right? And then there's some, like, very narrow, narrow other use cases like customer support, et cetera.”
Emberikos: Enterprise AI will consolidate around multi-functional super assistants
“But I think that's probably temporary, and then over time, I think we're gonna end up with agents that kind of can do anything for you. This is kind of what I was saying earlier. Like, there's just, like, a super assistant. You talk to it about anything.”
Stebbings: AI model providers will displace customer support software startups
“I do think, sorry, I do think, like, I think you're going to come for customer support, and I wouldn't want to be in that category.”
Emberikos: VCs must pivot from product builders to distribution-first founders
“I think this maybe changes what kind of founder you invest in, right? Like, I think there was this maybe temporary phase where, that I liked personally as a product builder. There was this phase where you would invest in, like, the person who can just, like, b…”
Emberikos: OpenAI struggles to close top candidates amidst fierce talent war
“You know, obviously at OpenAI, we have an incredibly strong brand, and so we're able to attract a lot of talent. But even so we put a ton of effort into, like, closing candidates that we're really excited about. Even like, even we feel it. It's not like, you d…”
Emberikos: Hiring an imperfect product manager does more harm than good
“You still need product people, but I do think that they have to be the perfect fit. And if you're, if you have someone who's like not the perfect fit, they're, they might just do more harm than good. So it's kind of means that like we're way more selective tha…”
Emberikos: As software creation gets easier, agency and taste become scarcer
“I think that because it's never been like easier to build things, the thing that becomes scarcer is like agency taste and like quality.”
Emberikos: Fastest agent productivity gains will come from desktop-level tools
“I do think that the fastest way we're going to see productivity gains from agents at work Is going to be at first meeting users on their computer, working with the stuff that they have available to them, you know, without having deployed FTEs to set anything u…”
Emberikos: Sourcegraph's AMP punches above its weight and initiated agent standards
“Their product has a great reputation of just being, like, you know, punching way above its weight, but I think the other thing that I really respect is that they helped initiate this whole, like, standardization around, like, agents.md and, like, dot agents sl…”
Emberikos: Software products cannot offer unlimited usage tiers for too long
“I think the lesson I learned the hard way there is like, you can't make things unlimited for too long.”
Emberikos: Startups will soon build on software stacks fully managed by AI
“I basically think that probably big companies will take a long time to, like, deploy this, but many startups might actually kind of start building on a completely new stack that's, like, fully AI managed.”
Emberikos: OpenAI is the only firm building OS-level sandboxing for Windows AI agents
“We're basically the only company that cares about OS level sandboxing for coding agents. For instance, there's none that exists on Windows. We're the ones building that and we're doing it in open source, so hopefully other people can use it.”