OpenAI is restructuring its core product stack around a single, unified agent platform. Thibault Sottiaux, recently promoted to head of core products, now oversees the merger of ChatGPT and Codex into what the company internally calls a "super app."
From a systems perspective, this is a consolidation play. OpenAI has already deprecated standalone products like Sora and its science platform, reallocating engineering resources toward a single runtime. Sottiaux, who built Codex from a research tool into one of OpenAI's fastest-growing revenue streams, now faces a different scaling problem: serving nearly a billion weekly active users with a generalized agent.
Architecturally, the super app is essentially Codex refactored into a general-purpose agent orchestration layer, then embedded into ChatGPT's existing infrastructure. Sottiaux describes the core challenge as converting Codex's developer-oriented toolchain into a system that executes tasks—writing code, calling APIs, scraping the web—without exposing the user to any of it. The model handles planning and execution; the user sees only natural language input and results.
OpenAI tried this before with Operator and ChatGPT Agent, but Sottiaux admits those launches were premature. The underlying models lacked reliability, forcing heavy guardrails that limited utility. Now, he claims the model quality has crossed a threshold where the system can be trusted with open-ended tasks.
One practical trade-off: user adoption. Engineers naturally understand how to prompt agents for complex workflows. Average users don't. Sottiaux's strategy is incremental releases, each building user confidence and teaching the model to act as a mentor. "You can't afford to do a big splash and be wrong," he notes.
The integration with existing infrastructure is another engineering challenge. Unlike WeChat, which built its own financial and messaging backbone, OpenAI must interface with existing systems like Gmail, Slack, and Visa's payment network. Recent partnerships suggest they're building a connector layer rather than reinventing those services.
Sottiaux's team remains lean—roughly 40 people two months ago—but has absorbed resources from shuttered projects. The goal is a single, personalized agent that handles everything from dinner reservations to expense reports, with long-term memory of user preferences and history. Whether the architecture can deliver that reliably at scale is the open question.
Source: Wired