OpenAI Exec Unpacks the Agent Architecture Behind 10M Codex and ChatGPT Work Users
Latent Space · rss · 2026-07-28
The latest Latent Space podcast featured Akshay Nathan, Head of Core Product Engineering at OpenAI, to deeply review the launch of ChatGPT Work and the product logic behind Codex's >10x user growth this year.
From Coding to Knowledge Work
OpenAI realized there are 100x more people who use code than those who can write it. As code generation becomes easier, non-developers represent the largest growth market. In June, knowledge workers already accounted for 20% of Codex's user base and were growing over 3x faster than developers. Consequently, OpenAI integrated the underlying agent architecture of Codex with ChatGPT Work.
Agent Architecture and Product Philosophy
- Unified Harness: Codex and ChatGPT Work share the same underlying agent engine, but feature different UX, Git visibility, and sandboxing defaults tailored to their respective audiences.
- Paradigm Shift: Traditional knowledge work is fragmented across documents, spreadsheets, and decks. ChatGPT Work allows users to simply describe an outcome, while the agent autonomously assembles tools, gathers context, and generates high-fidelity artifacts (e.g., replacing traditional decks with interactive websites).
- Context & Memory: Agents can gather context across code, Slack, documents, and local files, introducing advanced features like persistent compute environments, scheduled tasks, and sub-agents.
Workplace Insights for the AI Era
Akshay believes AI will turn more people into generalists with deep specialties. As the barrier to building software disappears, the real bottlenecks become personal "taste" and ideas. He also cautioned teams to distinguish between increased AI-generated motion and actual meaningful progress.
More from coding & agent
- CompAI says an AI browser agent can now log in and collect audit evidence — JosephJacks_ · 2026-07-29
- Yangqing Jai Teases New Venture: An AI Oracle to Mass-Produce Systems — jiayq · 2026-07-29
- The missing layer is not code generation, but autonomous software design and upkeep — jiayq · 2026-07-29
- AI could make custom software far more common and change software economics — jiayq · 2026-07-29
- Models write code fast, but production-ready software still needs another layer — jiayq · 2026-07-29
- An autonomous coding system built a database engine and passed 6 million SQLite tests — jiayq · 2026-07-29