Why humans are AI’s biggest bottleneck (and what’s coming in 2026) | Alexander Embiricos (OpenAI Codex Product Lead)

Codex is presented as the beginning of a software engineering teammate rather than an autocomplete tool. It can write code, use a computer, and eventually participate in planning, testing, deployment, and maintenance. The Sora Android app reached employees in 18 days and the public ten days later. Embiricos says the long-term goal is a proactive system that helps by default, while human typing, prompting, context configuration, and review remain important bottlenecks. OpenAI's fast, bottom-up experimentation depends on unusually strong autonomy and talent, so the approach cannot simply be copied by any company.

Codex is described as a coding agent that can be installed in an IDE or terminal and paired with a developer across the software lifecycle. The intended future is a proactive teammate that understands context, delegates work, and eventually helps with planning, validation, deployment, and maintenance. Growth accelerated after Codex moved from a harder-to-adopt cloud workflow into an interactive local IDE and CLI experience with a sandbox. The team combines product and research iteration, uses internal dogfooding while accounting for differences between OpenAI employees and the broader market, and treats human agency as a core product concern. The remaining bottleneck is often review and validation rather than code generation, which is why the team focuses on helping people trust and verify agent-written code.

Embiricos argues that the useful unit of progress is no longer only the model. OpenAI's Codex-Max combines a reasoning model with an API and a harness that determines how the model operates. He points to compaction as an example: a model that runs overnight or for a day can exceed its context window, so the model must understand when to prepare for a new context, the API must support that transition, and the harness must package the next payload. Shipping the feature required all three layers to be developed and tuned together.