OpenAI Codex lead on the new shape of product work | Andrew Ambrosino
Andrew Ambrosino, Product and Engineering Lead for the Codex app at OpenAI, details how AI is transforming product development, shifting focus from implementation to curation and the crucial role of "taste." He discusses AI's current limitations in design, the evolving nature of product roles, and OpenAI's "zone defense" approach to product management. Ambrosino shares his personal workflow with Codex and outlines the vision for it as an intelligent "home base" that orchestrates work across various applications, exemplified by a story of Codex building a Premiere Pro extension.
Andrew Ambrosino, Product and Engineering Lead for the Codex app at OpenAI, offers a deep dive into how AI is fundamentally reshaping product development and organizational structures. He emphasizes that AI has flipped the product development process, making implementation cheap. Now, anyone can build anything, shifting the challenge from coding to curation and discerning what is truly valuable amidst abundant prototypes.
The Ascendance of “Taste” and Curation
Ambrosino highlights that “taste” has emerged as the most valuable capability in an AI-first workplace. He clarifies that taste extends beyond aesthetics to encompass strategic judgment: knowing “what to work on, how to present that information, how to achieve the goals, what medium to use.” This includes systems thinking and understanding wider context. He cites Linear’s Head of Product and Paul Graham to illustrate that taste is about good judgment and discerning signal from noise, rather than mere aesthetic preference.
AI’s Limitations in Design
Despite AI’s rapid advancements, Ambrosino notes that current frontier models are “still bad at design.” This is attributed to design being harder to grade than code, as it involves the human aspect of taste as a feedback mechanism. Furthermore, design requires novelty and an element of randomness, which differs from software engineering where predictable patterns are often preferred. The deeper abstraction layer in design, connecting visual elements to underlying code structure and ensuring adaptability for rebranding, remains a significant challenge for AI.
The Evolving Design Process and Role of Documentation
The notion that “PRDs are dead, prototypes are in” is a simplification. While implementation is cheap, documents are still crucial for achieving clarity in vague product areas, while prototypes excel at stress-testing interaction patterns. The key is choosing the right medium for the message. He warns against the “primal mark” – over-anchoring on a polished prototype that was only meant for exploration. The process itself isn’t dead, but its rigid historical tools and stages are; understanding where you are in the process remains vital.
OpenAI’s “Zone Defense” for Product Management and Team Structure
OpenAI operates with a unique “zone defense” model for product management. With everyone being highly agentic and capable of building, product managers focus on identifying gaps, steering, and aligning the numerous explorations happening across the company. This ensures broad coverage while allowing high-agency individuals to drive initiatives. Ambrosino notes a significant “role collapse” within the Codex team, where traditional boundaries between engineering, design, and product management blur. Roles are defined by the “average of where they’re working,” with individuals often possessing multidisciplinary skills (e.g., designers who code, PMs with computer science backgrounds). However, he cautions against completely eliminating roles, as it risks discarding valuable best practices and specialized disciplines built over years.
Planning and Ambition in an AI-First World
Traditional roadmap planning is challenging due to the rapid pace of model advancements. Long-term precision is “false precision.” OpenAI’s approach involves continuously prototyping all interesting ideas, then selecting what’s ready now and allowing others to “bake” until models catch up. He gives the example of the Codex app’s success: if launched just a few months earlier, before critical model improvements, it “would have absolutely failed.” This highlights the need for both ambition and an understanding of market and model readiness, avoiding being “AGI-pilled” by releasing features too early.
Andrew’s Personal Workflow and the Vision for Codex
Ambrosino extensively uses Codex to manage his own evolving job role, from coding to product discovery and leadership. He automates daily briefs, monitors Slack channels, updates status trackers, and conducts internal research, effectively “coaching” the AI to refine its outputs. This personal workflow serves as a form of product discovery, identifying common automation patterns that can be generalized into first-class features (e.g., memory functions).
The overarching vision for Codex is to become the ultimate “home base” for knowledge workers. This includes integrating seamlessly with existing tools like Excel or Premiere Pro, not by replacing them, but by orchestrating tasks through connectors, computer use, or custom extensions. He shares the remarkable story of an OpenAI videographer using Codex to edit videos, where Codex even built its own extension for Premiere Pro to achieve tasks it couldn’t natively perform. This demonstrates the potential for Codex to act as an intelligent orchestrator, using whatever tools are necessary to complete a task, providing a cohesive and powerful user experience.