Why humans are AI's biggest bottleneck (and what's coming in 2026) | Alexander Embiricos (OpenAI)
Alexander Embiricos, product lead for OpenAI's Codex, shares the vision of AI as a software engineering teammate, not just a tool. He explains how a strategic shift to a local, interactive experience unlocked 20x growth, details how the Sora Android app was built in 28 days, and argues that the real bottleneck to AGI-level productivity is now human review speed, not model capability.
The Vision: From Coding Tool to AI Teammate
Codex is envisioned as the beginning of a “software engineering teammate,” not merely a tool. Today, it operates like a brilliant intern who has powerful capabilities but lacks broader context—it “refuses to read Slack” and “doesn’t check DataDog unless you ask it to.” The ultimate goal is for Codex to be a proactive agent that participates across the entire software development lifecycle, from ideation and planning to validation, deployment, and maintenance.
A core objective is to achieve proactivity, moving beyond a model that only responds to prompts. For an AI to be a true super-assistant, it must be able to do things. The team learned that the most effective way for models to interact with the world is by using a computer, and the best way for a model to use a computer is to write code. This leads to a powerful conclusion: to build any effective agent, you should start by building a coding agent. Coding is becoming the core competency for any AI agent that needs to perform tasks.
Unlocking Explosive 20x Growth
Codex has seen explosive growth, scaling more than 20x since August and serving trillions of tokens weekly. A key product decision unlocked this trajectory. The initial version of Codex was a cloud-based agent that users could delegate tasks to asynchronously. While powerful and heavily used within OpenAI, it was too far ahead for the general market. It required significant setup and a different way of thinking that was hard to adopt initially.
The pivotal shift was to meet developers where they are. By launching as an IDE extension and CLI tool that works interactively on a user’s local machine, Codex provided immediate, intuitive value. This approach allows users to build trust and a strong feedback loop with the agent. As users work with Codex side-by-side, they naturally configure it with the context and permissions it needs, creating a seamless on-ramp to the more powerful, delegated workflows that represent the future.
The Shifting Bottleneck: From Writing to Reviewing
As AI agents drastically accelerate the ability to write code, the primary bottleneck in software development is shifting. Historically, writing code was the time-consuming part. Now, the challenge is the cognitive load of reviewing and validating large amounts of AI-generated code. For many engineers, this is a less enjoyable and more taxing part of the job.
Recognizing this shift, the Codex team is focusing on features that address the review process. The goal is to build user confidence and make validation easier by:
- Developing features specifically for code review.
- Enabling the agent to better validate its own work by running tests and checks.
- Designing user experiences that prioritize clarity and trust, such as showing a visual preview of a UI change before presenting the code diff.
Real-World Impact and New Workflows
The acceleration enabled by Codex is already transforming how software is built at OpenAI.
- The Sora Android App: A completely new application was built from zero to an internal launch in just 18 days, and to a public launch 10 days later (28 days total). This was accomplished by a small team of two or three engineers. Codex excelled at porting logic from the existing iOS application, dramatically speeding up the process.
- The Atlas Browser: Building a web browser is a monumental task. The Atlas team reported that complex features that would have previously taken 2-3 engineers several weeks can now be completed by one engineer in a single week.
- Internal Acceleration: Codex is used to write the code that manages its own training runs and key infrastructure. The team is even experimenting with having Codex be “on call” for its own training, where it monitors performance charts and can potentially fix issues autonomously.
This power is also “compressing the talent stack,” blurring the lines between roles. Designers and product managers now write “throwaway code” to build interactive data viewers for analysis or create bespoke tools, like an animation editor, to solve a specific problem. Prototyping is often faster than writing a spec.
The Human Factor and the Path to AGI
The most underappreciated limiting factor to achieving AGI-level productivity is not model capability, but “literally human typing speed” and the time required for manual validation. The next major leap in productivity will come from unblocking these human-in-the-loop dependencies, allowing agents to become “default useful” without constant prompting and review.
This transition will be gradual. Startups building on modern tech stacks will be the first to experience this “hockey stick” growth in productivity. Over time, larger enterprises will adapt their complex, legacy systems to allow agents to handle more work end-to-end. The acceleration this provides will then feed back into AI research labs, creating a virtuous cycle that pushes toward AGI.