Why humans are AI's biggest bottleneck (and what's coming in 2026) | Alexander Embiricos (OpenAI)
Alexander Embiricos, product lead for OpenAI's Codex, discusses the vision of AI as a proactive software engineering teammate, the product decisions that led to its explosive 20x growth, and why the real bottleneck to AGI-level productivity is shifting from model capability to human review speed.
The Vision for Codex: From Tool to Teammate
Codex is envisioned as more than just a coding agent; it’s the beginning of a “software engineering teammate.” Today, it operates like a brilliant intern who is incredibly focused but lacks broader context—it “refuses to read Slack” and “doesn’t check DataDog unless you ask it to.” The primary goal is to evolve Codex toward proactivity, enabling it to participate across the entire software development lifecycle, from ideation and planning to validation, deployment, and maintenance.
The ultimate aim is to create a “super assistant” that is helpful by default, reducing the cognitive load on users who currently must be thoughtful about when and how to prompt an AI. If an AI can only help when prompted, its utility is limited. A true teammate, however, understands context and can offer assistance proactively, potentially thousands of times a day.
Unlocking Explosive Growth
Codex has seen explosive growth, scaling 20x since August and serving trillions of tokens weekly. A key product unlock was shifting the initial user experience from a powerful but complex cloud-based agent to a more intuitive, interactive IDE extension and CLI tool.
- Meeting Users Where They Are: The initial cloud-native version was “too far in the future.” It required users to configure environments and learn a new, asynchronous way of working. The breakthrough came from creating a local, interactive experience where the agent works alongside the developer on their machine. This provided an immediate, low-friction on-ramp to Codex’s capabilities.
- Integrated Development: The Codex team tightly integrates product and research, allowing for rapid iteration on the model, API, and harness simultaneously. This holistic approach enables features like “compaction,” which allows the model to work on tasks for extended periods (e.g., overnight) by managing its own context window—a feature requiring innovation at all three layers of the stack.
- Optimizing for a Specific Harness: By focusing the model’s training on a specific harness—in this case, using the shell within a secure sandbox—the team can move much faster than if they had to optimize for every possible tool and workflow.
The Shifting Bottleneck: From Writing to Reviewing
As AI agents become proficient at writing code, the human bottleneck in software development is shifting.
- Writing vs. Reviewing: Writing code is often a creative and enjoyable part of engineering. With agents generating vast amounts of code, engineers now spend more time on the less enjoyable task of reviewing AI-generated code.
- Product Focus on Acceleration: The product team’s goal is to “maximally accelerate people” rather than creating tools that make the human’s role unclear. This means focusing on features that build confidence in the AI’s output, improve the code review experience, and empower the human developer. For example, when an agent makes a UI change, showing a visual preview first is more empowering than immediately presenting a code diff.
The Future of Agents and Development Workflows
Coding is emerging as a core competency for any effective AI agent. To perform tasks and interact with the world, an agent needs to use a computer, and the most effective way for a model to use a computer is by writing code. This suggests that any future general-purpose agent will likely be a coding agent at its core.
Development workflows are also evolving beyond simple “prompt-to-patch” interactions.
- Plan-Driven Development: For complex, long-running tasks, a successful pattern is to first collaborate with Codex to generate a detailed plan (e.g., a markdown file). Once the human approves the plan, the agent can execute it more reliably.
- Chatter-Driven Development: The future may not be spec-driven but “chatter-driven,” where an agent acts on ambient signals from team communications, customer feedback channels, and social media to proactively identify, fix, and deploy solutions for small bugs without requiring a formal spec.
Building at OpenAI Speed: Case Studies
The internal adoption of Codex at OpenAI demonstrates its power to accelerate development.
- Sora Android App: The team built and shipped a completely new, number-one-ranked app in just 28 days (18 days to an internal version, 10 more to public launch) with only a handful of engineers. Codex was particularly effective at porting logic by looking at the existing iOS application and the target Android codebase simultaneously.
- Atlas Browser: On the Atlas browser project, Codex provided a massive productivity boost, turning tasks that would have previously taken 2-3 engineers for 2-3 weeks into tasks that one engineer could complete in a week.
The Human Bottleneck and the Path to AGI
The most underappreciated limiting factor in achieving AGI-level productivity isn’t model capability but “literally human typing speed or human multitasking speed.” The real bottleneck is the human-in-the-loop for prompting the AI and manually validating all of its work.
The “hockey stick” of productivity growth will be unlocked when systems are rebuilt to allow agents to be default useful and more autonomous in validating their own work. This won’t be a binary event but a gradual process. Startups will be the first to adopt new stacks that enable this autonomy, with larger enterprises following as they slowly update their complex legacy systems. This acceleration will eventually feed back into the AI labs themselves, creating the flywheel that leads toward AGI.