Codex from 0 to 10M Users: Building ChatGPT Work — Akshay Nathan, OpenAI

What It Covers

Akshay Nathan, OpenAI's core product engineering lead, explains how ChatGPT Work reached 10 million users by merging the Codex and ChatGPT harnesses into a unified productivity platform. The conversation covers the product architecture decisions, sub-agent design tradeoffs, memory systems, artifact evolution, and how OpenAI measures team progress in an AI-accelerated development environment.

Key Questions Answered

• •Harness unification: ChatGPT Work and Codex share the same underlying capability harness — the differences are purely UX-level. Codex surfaces git diffs and file edits explicitly; Work abstracts those details away. Users never need to choose between them because the model routes automatically based on task type, such as redirecting spreadsheet creation toward Work mode without manual switching.
• •Non-developer adoption signal: When Codex launched internally at OpenAI, the most telling signal was not developer usage but the pride non-technical staff — strategic finance, marketing — felt using it. That emotional response, feeling like they had a superpower, became the product thesis for ChatGPT Work: agents are ready for general knowledge workers, not just engineers, earlier than anticipated.
• •Motion vs. progress trap: Teams adopting AI tools frequently confuse activity with advancement. Nathan distinguishes motion — easy to generate with current tooling — from progress, which requires deliberate, prescriptive goal-setting upfront. Managers should define what a completed "at bat" looks like: idea generation, build, user feedback, hypothesis validation, and iteration, all completed efficiently in sequence.
• •Sub-agent defaults: Ultra mode, which activates multi-agent parallelization, was moved behind an advanced settings toggle after launch feedback showed casual users were hitting usage limits unexpectedly. The practical guidance: use Ultra for tasks that are either highly parallelizable or genuinely open-ended; use default configuration for most tasks and adjust only when quality or cost efficiency falls short.
• •Sites replacing slide decks: OpenAI's internal corporate finance team now produces monthly reports as hosted sites rather than PowerPoint or Excel files. Because sites are HTML-based, they are infinitely flexible — any layout, any data prominence — without requiring the user to know a specific feature. Nathan identifies this format shift as an underreported consequence of the ChatGPT Work launch alongside artifacts.
• •Memory as retention flywheel: ChatGPT Work inherits the memory v3 system by default, meaning all Work sessions read from and write back to a user's existing ChatGPT memory profile. The practical implication: the longer someone has used ChatGPT conversationally, the richer the contextual foundation Work agents draw from immediately, without any manual setup or re-onboarding when switching to agentic tasks.