OpenAI's Codex Lead: Why Coding as We Know It is Over

Summary

• Coding “automation” means more engineers, not fewer. Alexander Embiricos (product lead for Codex at OpenAI) agrees with Elon that coding is the first domain where LLMs are really good, but rejects the framing: when assembly gave way to higher-level languages, “did we say coding is automated? Not really” — demand for code exploded and headcount grew. His call: more builders in five years, with the talent stack compressing into full-stack generalists and PMs often useful but needed less until teams are large (“I don’t think you need them”).
• Alex’s earlier claim was that human typing speed and validation work are a key bottleneck to AGI, not model compute or architecture; he also says there are multiple bottlenecks. Harry uses AI 30+ times a day; Alex thinks it “should be helping us tens of thousands of times per day, compute budget permitting” — and admits even he is “too lazy to type out that many prompts and too uncreative to figure out all the ways AI can help me.” The endgame is AI that “knows you, connected to your context and chimes in helpfully” with zero prompting effort.
• “All agents are actually coding agents because coding is just the best way for an agent to use a computer.” His three phases — coding agents, then open-ended computer-use tools for tinkerers, then out-of-the-box productization — will be “speedrun… in the next months.” Builders are already using the Codex app for non-coding tasks.
• Inside OpenAI, most people Alex knows are basically not opening editors. The step change was GPT 5.2 Codex in December — full delegation (“do a plan with it… then I’m just going to go let it cook”), the vast majority of code AI-written, and nearly all code auto-reviewed by Codex. Codegen “has become basically trivial”; the underinvested bottleneck is now review and trust — the goal is an agent you trust to own an entire microservice with no human review.
• Strategy is open now, sticky later — and margins can wait. Coding tasks are “hermetic… episodic” and vendor-neutral today (AGENTS.md, open-source harness), but connecting agents to Sentry or Google Docs “is a sticky decision” enterprises won’t repeat. “I view this year as a race” — you should be “okay taking some hit to margin” to win the connection layer.
• “Our job is the distribution of intelligence” — OpenAI serves its models to competitors and Alex says that’s rational long-game play, while Codex’s real advantages are ChatGPT distribution plus training its own models for its harness and building the harness for new models. North star is active users, not revenue; growth since August was “like 20x,” with a further increase from December to then, followed by the Codex app, a Super Bowl ad and free ChatGPT-tier access.
• Terminal market: “fewer providers capturing a lot of value” — one super assistant becomes the Slack-style center of gravity of work, not 12 specialized agents. For investors: SaaS that owns a human relationship or a system of record survives (“maybe more important than ever”); glue layers are in danger; Harry adds physical infrastructure and gnarly fintech are safe — and customer support is not.

Deep dive

1. Automation means more engineers — and a compressed talent stack

• Asked whether Elon is right that coding will be among the first professions largely automated, Alex reframes: coding is the first domain where LLMs are really good, but “automated” is “kind of a heavy statement.” His analogy, as told: when we stopped writing assembly, “did we say coding is automated? Not really… we were just able to write much more code and then as a result actually there was much more demand for code and there were many more software engineers required.” Same with “computer” — originally humans at Bletchley Park punching cards; every time the task was automated, “there’s been an explosion in demand for the output.”
• So: more engineers in five years, not less — but the talent stack compresses. The backend/front-end split is already fading on the Codex team; “engineer” will mean someone far more full-stack. His running joke that you don’t need PMs is qualified: the role is “explicitly undefined,” and everything he does “could be done by a really strong eng lead or a designer” — though he later clarifies that you still need product people, but only perfect fits, since a bad one “might just do more harm than good.”
• His advice to a CS student: “there’s actually never been a better time to be an engineer.” What becomes scarce as building gets easy is “agency, taste and quality” — build things, share them; “an interesting project gets my attention much more than a normal resume does.” Meanwhile the war for talent is “incredibly fierce” — even OpenAI puts “a ton of effort into closing candidates.”

2. A key AGI bottleneck: human effort

• Alex runs it Socratically: Harry uses AI 30+ times a day; Codex usage sits in the “tens of times” range; “I think AI should be helping us tens of thousands of times per day… compute budget permitting.” The gap is human effort — “I’m too lazy to type out that many prompts and I am too uncreative to figure out all the ways that AI can help me.”
• The power-user pattern shows what’s coming: OpenAI engineers tell him “I constantly have Codex running. I never close my laptop and if it’s not running while I’m in a meeting I’m like wasting my time.” But keeping agents fed is itself work — that management overhead is the bottleneck.
• The destination: AI you don’t need to prompt at all — “it just knows you, connected to your context and chimes in helpfully.” Most people shouldn’t need effort or creativity to benefit from AGI.

3. Three phases of agents — and all agents are coding agents

• The roadmap in his head: first, agents that work really well for software engineering “because LLMs happen to be good at that.” Second, realize “all agents are actually coding agents because coding is just the best way for an agent to use a computer” — take the same flexible tool to anyone excited to tinker (builders already use the Codex app for triaging tasks, monitoring deploys, other non-coding work). Third, productize what proves out into features that “just work immediately out of the box.” His timeline: “we’re going to speedrun this entire 1-2-3 journey in the next months.”
• The design lesson he credits to a rival: “the genius of when Claude Code first shipped” was an open-ended tool, easy to use in any context, right in the terminal — people discovered the use cases themselves. The mistake to avoid is over-verticalizing (“AI capabilities but only specifically for finance”) before you know the right form factor.

4. The FDE fight: bottom-up beats top-down in the enterprise

• Matt Fitzpatrick (Invisible AI) said on a prior 20VC that you can’t adopt AI at enterprise without forward-deployed engineers. Alex: “even though I am literally hiring FDEs… I actually disagree with that entirely.” Top-down workflow automation “massively underleverages” AI; instead give it to the people doing the work so they build intuition and pull automation into their own workflows — empowering, versus automation arriving as a complete, disempowering intervention.
• Harry’s pushback, worth keeping: data security, permissioning and access provisioning are “really freaking hard,” and enterprise users are “much less intelligent and confident than we give them credit for” — you need an FDE to custom-fit horizontal tools. Alex concedes the point for grand zero-to-one workflow systems, but his counter: every workflow terminates at an employee’s browser or file system — “everything comes to an interface that an agent running locally on your computer can work with.”
• That’s a key reason OpenAI is building the Atlas browser: by controlling it tightly end to end, they can offer “safe agentic browsing for enterprise” — agentic access to systems FDEs haven’t wired up yet.

5. Inside OpenAI: nobody opens an editor anymore

• The step function was GPT 5.2 Codex in December: before it, AI coding meant tab completion or pair programming with hands on the keyboard; after it, “I’m just going to fully delegate this task… do a plan with it, make sure we like the spec… and then I’m just going to go let it cook.” Today “most people that I know are basically not opening editors anymore,” the vast majority of code is AI-written, and “the code itself is not being written by humans anymore” — humans keep the module interfaces and the plans. The Codex app deliberately has no text editing, so it’s clear you’re delegating, not pairing.
• Review moves up the stack. Plan review — the agent proposes a long plan and asks questions, like a new hire presenting an RFC — “is actually something that’s becoming more important.” For actual code review, Codex is explicitly trained for it with “few false positives of criticism,” and “nearly all code at OpenAI is reviewed by Codex automatically.” The funnel trick: people ask Codex to review a different model’s code and conclude “oh shoot, I should probably just be using Codex.”
• Speed is “super important” but no inference monopoly will form despite the Cerebras partnership (“news coming… soon”) — competitive pressure means he expects “multiple answers.” Meanwhile GPT 5.3 Codex is significantly more efficient, and inference changes made API serving ~40% faster and Codex ~25% faster.
• What he’d invest in differently: “writing code has become basically trivial,” but review and trust bottlenecks are “underappreciated still and underinvested in.” The goal is an “unbottlenecked” agent trusted to own an entire microservice and run the full iterative loop, user feedback included, without human review — hard on intelligence and equally hard on safety and controls.

6. Retention: open now, sticky later — this year is a race

• Codex is built “super openly”: the core harness is open source, and OpenAI deliberately named the instruction file AGENTS.md rather than CODEX.md — “pretty much every agent except Claude uses AGENTS.md.” Same story with skills stored in a neutral .agents folder: “everyone has jumped on it except the usual suspect.”
• Why give up lock-in? Coding tasks are “quite hermetic… episodic” — an AGENTS.md file in, a git patch out, both ends vendor-neutral, so switching is easy today anyway. Stickiness arrives when agents start touching Sentry, Google Docs, systems of record: “deciding to connect an agent to that system is a sticky decision” enterprises won’t make twice — which is why OpenAI runs “the most conservative sandboxing approach” and is “basically the only company that cares about OS-level sandboxing,” building the Windows version in open source.
• On margins, his answer to Harry’s software-margin worry: costs will come down significantly, but “if this is the year of agents being deployed broadly at work… I view this year as a race. You want to win that race and you should be okay taking some hit to margin in the meantime.”

7. “Our job is the distribution of intelligence” — and how Codex wins

• The bit Harry says is “so difficult for me as a venture capitalist to understand”: OpenAI trains frontier models, then serves them to competitors. Alex’s logic: “we’re playing such a long game… if the competition gets better, we learn.” He’d quote-tweeted a Warp launch that very morning — “we’re all kind of inevitably reaching the same conclusions together.” Codex’s real edges: ChatGPT distribution and training its own models to be good in its harness, while building the harness to be good at new models, with early access nobody else gets.
• Pushed on the defining factor of winning, the company answer (“above my pay grade”) is compute advantage and having the best models — successful products exist partly to generate revenue and to pressure-test the model faster. The product answer: build for individuals, let them become fluent, then let them pull in automation — plus real enterprise GTM, since “feel free to use the stuff” doesn’t work without education and configuration.
• North star is active users, not revenue — weekly actives, specifically. Harry’s challenge: if you’re replacing the IDE, shouldn’t it be daily? Alex concedes: “daily active will be better soon… I actually agree with the criticism.” The staircase: Google for navigation, ChatGPT for information, and this year “for any task I need to do, I go to this text box and something happens that helps me.”
• Momentum is real and felt: the original cloud-agent launch “didn’t work as well as what we shipped later”; since August’s GPT-5 interactive push, “we grew by like 20x,” with what he described as a further doubling from December to then, though he forgot the exact number. GPT 5.3 Codex “cemented” the most-intelligent-model claim — a competitor model that launched 20 minutes earlier “was SOTA for 20 minutes… maybe this is spicy.” Then the app, the Super Bowl ad (“you can just build things”) and Codex on free ChatGPT plans. His other regret: it’s time to “get back to cloud,” now that users are fluent enough for the step up.

8. End state: one super assistant, with a Slack-style center of gravity

• Will chat endure as the UI? Yes — “sci-fi is a really good predictor of what the future should look like… simple is usually right”: one entity you talk to about anything, not separate coding and sales AIs. But power users pair it with bespoke GUIs — an executive assistant you can only talk to is “super annoying”; sometimes you want to edit the show notes yourself. A marketer chats; a builder goes deep in the Codex app.
• Market structure gets concentrated: “fewer providers capturing a lot of value.” Coding-only PMF is “probably temporary”; once agents generalize, a company with 12 agents means employees won’t achieve fluency. His Dropbox-era analogy, as told: commenting on the document in Dropbox was objectively “more optimal,” but “nobody wants to comment on the document. I just want to Slack you” — a single agent for nearly anything creates the same giant pull, becomes “the center of gravity of work,” and you end up with “just a handful of these.”
• On agent-to-agent design: the best interfaces for Codex “tend to be the best interfaces for humans” — his test-runner example: emitting only failed tests instead of hundreds of thousands of output lines is better for both. So agent-to-agent touchpoints will mirror human-in-the-loop ones, letting you “atomically replace individual systems.”

9. What survives the model providers — and who to back now

• His SaaS survival test: does the company own a relationship with a human, or a system of record? If either, “it’s probably not going away — maybe both of those are more important than ever.” A glue layer owning neither: “I’m not the expert here, but I’m more nervous about that kind of company.” Harry’s read: the SaaS selloff is “massively exaggerated” — Monday.com is fine because vibe-coding a to-do list isn’t cost-efficient — but “respectfully, I think Dropbox is in a very difficult position,” and “I think you’re going to come for customer support and I wouldn’t want to be in that category.”
• The founder profile shifts too: the era of backing pure product-builders while ignoring their GTM thesis “was an anomaly.” Now that building good product is relatively easier, back the founder with distribution thinking and domain expertise for a specific customer.
• Where won’t model providers go? Alex’s honest non-answer: “I’m tempted to just say I don’t know… it’s a hard time to be an investor.” Harry supplies his own: physical infrastructure and energy supply, plus gnarly fintech — “I don’t think OpenAI is going to go into building 500 relationships with banks in Southeast Asia.” Alex tends to agree: gnarly markets where customer relationships and market knowledge are everything “still seem great.”
• On the coding data moat (a question relayed from an unnamed VC — “he who shall not be named”): “we feel like we have plenty enough data to build really good coding models.” The scarce data is knowledge-work trajectories, not code — maybe pay people to simulate tasks, “maybe you should acquire startups… that are no longer in business… say their Slack.” For scaled data campaigns, OpenAI enlists external data providers rather than building in-house — “how do we move as quickly as possible.”

10. Quickfire: wrong on multimodal, burned by unlimited pricing, aiming for grandma

• Biggest change of mind: when he joined OpenAI he expected we’d be screen-sharing and talking to agents within a year — “that was completely wrong.” Multimodal progress was slower than expected; instead, “agents that work with your computer through code are the way” — a complete rethink of how AI’s benefits reach ordinary people.
• Most painful product decision: Codex Cloud was effectively unlimited, and every day it stayed that way made winding it back harder. When they finally did, blowback from “a very small minority of users” polluted social chatter everywhere. Lesson learned the hard way: “you can’t make things unlimited for too long.”
• Most respected lesser-known competitor: Amp, out of Sourcegraph — a product “punching way above its weight,” and credit for kicking off AGENTS.md standardization via Quinn’s tweet offering to adopt whoever bought the domain. On Anthropic’s ads, the response he endorses: “one company’s being pretty negative about the future and the other company, us, OpenAI, is being really positive and just telling people they can build things and to dream.”
• In five years we’ll be embarrassed by editing code by hand — and, “maybe spicier,” by hand-managing deploys and monitoring: startups will build on a fully AI-managed stack (which “doesn’t exist yet”), where you start a company by getting an agent and asking it to build, then add co-founders to your agent tool. The ten-year vision is closing the capability overhang for non-tech users: “at some point we’ll add an agent to our family WhatsApp and it’ll just start being useful to the family” — especially for people like his grandma.